# Import PyDrive and associated libraries.
# This only needs to be done once per notebook.
from pydrive.auth import GoogleAuth
from pydrive.drive import GoogleDrive
from google.colab import auth
from oauth2client.client import GoogleCredentials

# Authenticate and create the PyDrive client.
# This only needs to be done once per notebook.
auth.authenticate_user()
gauth = GoogleAuth()
gauth.credentials = GoogleCredentials.get_application_default()
drive = GoogleDrive(gauth)

# List .txt files in the root.
#
# Search query reference:
# https://developers.google.com/drive/v2/web/search-parameters
listed = drive.ListFile({'q': "title contains '.csv' and 'root' in parents"}).GetList()
for file in listed:
  print('title {}, id {}'.format(file['title'], file['id']))

WARNING:root:pydrive is deprecated and no longer maintained. We recommend that you migrate your projects to pydrive2, the maintained fork of pydrive

title train.csv, id 1Jwzj9lneJfyHnmDgw75FpN8WFVwlgrQG
title test.csv, id 1GblmLmA4sm5bi1X3NmaPqNpDgnT3M3bZ
title cyberlab.csv, id 1TSw0JJGfrlzrgvlgFK21QBjAzcL_bzRD
title dataset_lab_1.csv, id 14rHwK3VAyCyVaSaz3ng2f8uitgFWr__9

# Download a file based on its file ID.
#
# siome hd
file_id = '14rHwK3VAyCyVaSaz3ng2f8uitgFWr__9'
# ale hd
#file_id = '1KLGXUxWYPTLA5HDtaEpKhVIXOkiSNV4I'
downloaded = drive.CreateFile({'id': file_id})
print('Downloaded content "{}"'.format(downloaded.GetContentString()))

from google.colab import drive
drive.mount('/content/drive')

Mounted at /content/drive

from torch.nn.utils.rnn import pad_sequence, pack_padded_sequence, pad_packed_sequence
from torch.utils.data import DataLoader, Dataset, TensorDataset, random_split
import torch.optim as optim
import torch.nn as nn
import torch

from sklearn.metrics import accuracy_score, confusion_matrix, ConfusionMatrixDisplay
from sklearn.preprocessing import StandardScaler, MinMaxScaler
from scipy.io.arff import loadarff
from tqdm import tqdm
import pandas as pd
import numpy as np
import time
from sklearn.model_selection import train_test_split


import matplotlib.pyplot as plt
import seaborn as sns

from pydrive.drive import GoogleDrive
from pydrive.auth import GoogleAuth
from google.colab import auth
from oauth2client.client import GoogleCredentials
# Authenticate and create the PyDrive client.
auth.authenticate_user()
gauth = GoogleAuth()
gauth.credentials = GoogleCredentials.get_application_default()
drive_pydrive = GoogleDrive(gauth) # use a different variable name here

downloaded = drive_pydrive.CreateFile({'id': file_id}) #call CreateFile on the pydrive object
downloaded.GetContentFile('data_set_lab_1.csv')  # Download to a local file
df = pd.read_csv('data_set_lab_1.csv')

df.head()

df.describe()

grouped_df = df.groupby('Label')
grouped_df.mean()

grouped_df.describe()

cat_cols = df.select_dtypes(include=['object', 'category']).columns.tolist()
print("Categoriche (object/category):", cat_cols)

Categoriche (object/category): ['Label']

from pandas.api.types import is_categorical_dtype

explicit_cats = [col for col in df.columns if is_categorical_dtype(df[col])]
print(explicit_cats)

categorical_features = list(set(explicit_cats))
print(categorical_features)

[]
[]

/tmp/ipython-input-11-3843240968.py:3: DeprecationWarning: is_categorical_dtype is deprecated and will be removed in a future version. Use isinstance(dtype, pd.CategoricalDtype) instead
  explicit_cats = [col for col in df.columns if is_categorical_dtype(df[col])]

#Now let's search for the rows whose features contain infinite values.
row_with_inf = df[df.isin([np.inf, -np.inf]).any(axis=1)]
print(row_with_inf.index)
print(row_with_inf)

Index([ 2130,  6683,  9265, 10138, 11793, 14833, 15627, 16361, 20043, 20524,
       20650, 20779, 20839, 21003, 21449, 21700, 22152, 22955, 23687, 24006,
       24045, 24404, 24452, 24454, 24598, 24608, 29324],
      dtype='int64')
       Flow Duration  Flow IAT Mean  Fwd PSH Flags  Bwd Packet Length Mean  \
2130               0            0.0              0                     0.0   
6683               0            0.0              0                     2.0   
9265               0            0.0              1                     0.0   
10138              0            0.0              0                     0.0   
11793              0            0.0              0                     0.0   
14833              0            0.0              0                     2.0   
15627              0            0.0              0                     6.0   
16361              0            0.0              0                     0.0   
20043              0            0.0              0                     0.0   
20524              0            0.0              0                     0.0   
20650              0            0.0              0                     0.0   
20779              0            0.0              0                     0.0   
20839              0            0.0              0                     0.0   
21003              0            0.0              0                     0.0   
21449              0            0.0              0                     0.0   
21700              0            0.0              0                     0.0   
22152              0            0.0              0                     0.0   
22955              0            0.0              0                     0.0   
23687              0            0.0              0                     0.0   
24006              0            0.0              0                     0.0   
24045              0            0.0              0                     0.0   
24404              0            0.0              0                     0.0   
24452              0            0.0              0                     0.0   
24454              0            0.0              0                     0.0   
24598              0            0.0              0                     0.0   
24608              0            0.0              0                     0.0   
29324              0            0.0              0                     6.0   

       Bwd Packet Length Max  Flow Bytes/s  Down/Up Ratio  SYN Flag Count  \
2130                       0           NaN              0               0   
6683                       2           inf              1               0   
9265                       0           inf              0               1   
10138                      0           inf              0               0   
11793                      0           inf              0               0   
14833                      2           inf              1               0   
15627                      6           inf              1               0   
16361                      0           NaN              0               0   
20043                      0           NaN              0               0   
20524                      0           NaN              0               0   
20650                      0           NaN              0               0   
20779                      0           NaN              0               0   
20839                      0           NaN              0               0   
21003                      0           NaN              0               0   
21449                      0           NaN              0               0   
21700                      0           NaN              0               0   
22152                      0           NaN              0               0   
22955                      0           NaN              0               0   
23687                      0           NaN              0               0   
24006                      0           NaN              0               0   
24045                      0           NaN              0               0   
24404                      0           NaN              0               0   
24452                      0           NaN              0               0   
24454                      0           NaN              0               0   
24598                      0           NaN              0               0   
24608                      0           NaN              0               0   
29324                      6           inf              1               0   

       Fwd Packet Length Mean  Fwd IAT Std  Packet Length Mean  \
2130                      0.0          0.0            0.000000   
6683                      6.0          0.0            4.666666   
9265                      7.0          0.0            9.333333   
10138                     6.0          0.0            6.000000   
11793                     6.0          0.0            6.000000   
14833                     6.0          0.0            4.666666   
15627                     6.0          0.0            6.000000   
16361                     0.0          0.0            0.000000   
20043                     0.0          0.0            0.000000   
20524                     0.0          0.0            0.000000   
20650                     0.0          0.0            0.000000   
20779                     0.0          0.0            0.000000   
20839                     0.0          0.0            0.000000   
21003                     0.0          0.0            0.000000   
21449                     0.0          0.0            0.000000   
21700                     0.0          0.0            0.000000   
22152                     0.0          0.0            0.000000   
22955                     0.0          0.0            0.000000   
23687                     0.0          0.0            0.000000   
24006                     0.0          0.0            0.000000   
24045                     0.0          0.0            0.000000   
24404                     0.0          0.0            0.000000   
24452                     0.0          0.0            0.000000   
24454                     0.0          0.0            0.000000   
24598                     0.0          0.0            0.000000   
24608                     0.0          0.0            0.000000   
29324                     0.0          0.0            2.000000   

       Fwd Packet Length Max  Subflow Fwd Packets  Flow Packets/s  \
2130                       0                    2             inf   
6683                       6                    1             inf   
9265                      14                    2             inf   
10138                      6                    2             inf   
11793                      6                    2             inf   
14833                      6                    1             inf   
15627                      6                    1             inf   
16361                      0                    2             inf   
20043                      0                    2             inf   
20524                      0                    2             inf   
20650                      0                    2             inf   
20779                      0                    2             inf   
20839                      0                    2             inf   
21003                      0                    2             inf   
21449                      0                    2             inf   
21700                      0                    2             inf   
22152                      0                    2             inf   
22955                      0                    2             inf   
23687                      0                    2             inf   
24006                      0                    2             inf   
24045                      0                    2             inf   
24404                      0                    2             inf   
24452                      0                    2             inf   
24454                      0                    2             inf   
24598                      0                    2             inf   
24608                      0                    2             inf   
29324                      0                    1             inf   

       Total Fwd Packets  Destination Port     Label  
2130                   2             52606    Benign  
6683                   1             64231    Benign  
9265                   2             47542    Benign  
10138                  2             17774    Benign  
11793                  2                21    Benign  
14833                  1             64299    Benign  
15627                  1               443    Benign  
16361                  2             19152    Benign  
20043                  2                80  DoS Hulk  
20524                  2                80  DoS Hulk  
20650                  2                80  DoS Hulk  
20779                  2                80  DoS Hulk  
20839                  2                80  DoS Hulk  
21003                  2                80  DoS Hulk  
21449                  2                80  DoS Hulk  
21700                  2                80  DoS Hulk  
22152                  2                80  DoS Hulk  
22955                  2                80  DoS Hulk  
23687                  2                80  DoS Hulk  
24006                  2                80  DoS Hulk  
24045                  2                80  DoS Hulk  
24404                  2                80  DoS Hulk  
24452                  2                80  DoS Hulk  
24454                  2                80  DoS Hulk  
24598                  2                80  DoS Hulk  
24608                  2                80  DoS Hulk  
29324                  1              1026  PortScan

# Checking null values
print(df.isnull().sum() , "\n---------------- \n"  )

Flow Duration              0
Flow IAT Mean              0
Fwd PSH Flags              0
Bwd Packet Length Mean     0
Bwd Packet Length Max      0
Flow Bytes/s              20
Down/Up Ratio              0
SYN Flag Count             0
Fwd Packet Length Mean     0
Fwd IAT Std                0
Packet Length Mean         0
Fwd Packet Length Max      0
Subflow Fwd Packets        0
Flow Packets/s             0
Total Fwd Packets          0
Destination Port           0
Label                      0
dtype: int64 
----------------

label_counts = df['Label'].value_counts()
print(label_counts)

# The dataset exhibits class imbalance, which may bias the model toward the majority class.
# We can handle it with resampling (e.g., SMOTE, ADASYN, undersampling) or algorithm-level techniques
# (e.g., class weight adjustments, focal loss). Also, we can use appropriate metrics to evaluate performance correctly.

Label
Benign         20000
DoS Hulk        5000
PortScan        5000
Brute Force     1507
Name: count, dtype: int64

from sklearn.preprocessing import LabelEncoder
label_encoder = LabelEncoder()
df['Label'] = label_encoder.fit_transform(df['Label'])
# we can also encode the port (we don't do that )
#df['Destination Port'] = label_encoder.fit_transform(df['Destination Port'])
df

# Regarding the features that should be " cleaned", from drescribe we see that we shoould look at "Flow bytes/s", "Flow Packets/s" and maybe "Fwd PSH Flags" and the ohters printed in the skewness anlysis

df.describe()

# Checking null values and then removing it
print(df.isnull().sum() , "\n---------------- \n"  )

Flow Duration              0
Flow IAT Mean              0
Fwd PSH Flags              0
Bwd Packet Length Mean     0
Bwd Packet Length Max      0
Flow Bytes/s              20
Down/Up Ratio              0
SYN Flag Count             0
Fwd Packet Length Mean     0
Fwd IAT Std                0
Packet Length Mean         0
Fwd Packet Length Max      0
Subflow Fwd Packets        0
Flow Packets/s             0
Total Fwd Packets          0
Destination Port           0
Label                      0
dtype: int64 
----------------

#Now let's search for the rows whose features contain NaN values.
rows_with_null = df[df.isnull().any(axis=1)]
rows_with_null

# Checking for all the rows with inf as fall
inf_val = df[df.isin([np.inf, -np.inf]).any(axis=1)]
inf_val

# 1. How many rows contain 'inf' values?
rows_with_inf = df[df.isin([np.inf, -np.inf]).any(axis=1)]
num_rows_with_inf = len(rows_with_inf)
print(f"Number of rows with infinite values: {num_rows_with_inf}")

# 2. rows with 'inf' values
print("\nRows with infinite values:")
print(rows_with_inf)

# 3. Which features contain 'inf' values?
columns_with_inf = df.columns[df.isin([np.inf, -np.inf]).any()]
print("\nFeatures with infinite values:")
print(columns_with_inf.tolist())

Number of rows with infinite values: 27

Rows with infinite values:
       Flow Duration  Flow IAT Mean  Fwd PSH Flags  Bwd Packet Length Mean  \
2130               0            0.0              0                     0.0   
6683               0            0.0              0                     2.0   
9265               0            0.0              1                     0.0   
10138              0            0.0              0                     0.0   
11793              0            0.0              0                     0.0   
14833              0            0.0              0                     2.0   
15627              0            0.0              0                     6.0   
16361              0            0.0              0                     0.0   
20043              0            0.0              0                     0.0   
20524              0            0.0              0                     0.0   
20650              0            0.0              0                     0.0   
20779              0            0.0              0                     0.0   
20839              0            0.0              0                     0.0   
21003              0            0.0              0                     0.0   
21449              0            0.0              0                     0.0   
21700              0            0.0              0                     0.0   
22152              0            0.0              0                     0.0   
22955              0            0.0              0                     0.0   
23687              0            0.0              0                     0.0   
24006              0            0.0              0                     0.0   
24045              0            0.0              0                     0.0   
24404              0            0.0              0                     0.0   
24452              0            0.0              0                     0.0   
24454              0            0.0              0                     0.0   
24598              0            0.0              0                     0.0   
24608              0            0.0              0                     0.0   
29324              0            0.0              0                     6.0   

       Bwd Packet Length Max  Flow Bytes/s  Down/Up Ratio  SYN Flag Count  \
2130                       0           NaN              0               0   
6683                       2           inf              1               0   
9265                       0           inf              0               1   
10138                      0           inf              0               0   
11793                      0           inf              0               0   
14833                      2           inf              1               0   
15627                      6           inf              1               0   
16361                      0           NaN              0               0   
20043                      0           NaN              0               0   
20524                      0           NaN              0               0   
20650                      0           NaN              0               0   
20779                      0           NaN              0               0   
20839                      0           NaN              0               0   
21003                      0           NaN              0               0   
21449                      0           NaN              0               0   
21700                      0           NaN              0               0   
22152                      0           NaN              0               0   
22955                      0           NaN              0               0   
23687                      0           NaN              0               0   
24006                      0           NaN              0               0   
24045                      0           NaN              0               0   
24404                      0           NaN              0               0   
24452                      0           NaN              0               0   
24454                      0           NaN              0               0   
24598                      0           NaN              0               0   
24608                      0           NaN              0               0   
29324                      6           inf              1               0   

       Fwd Packet Length Mean  Fwd IAT Std  Packet Length Mean  \
2130                      0.0          0.0            0.000000   
6683                      6.0          0.0            4.666666   
9265                      7.0          0.0            9.333333   
10138                     6.0          0.0            6.000000   
11793                     6.0          0.0            6.000000   
14833                     6.0          0.0            4.666666   
15627                     6.0          0.0            6.000000   
16361                     0.0          0.0            0.000000   
20043                     0.0          0.0            0.000000   
20524                     0.0          0.0            0.000000   
20650                     0.0          0.0            0.000000   
20779                     0.0          0.0            0.000000   
20839                     0.0          0.0            0.000000   
21003                     0.0          0.0            0.000000   
21449                     0.0          0.0            0.000000   
21700                     0.0          0.0            0.000000   
22152                     0.0          0.0            0.000000   
22955                     0.0          0.0            0.000000   
23687                     0.0          0.0            0.000000   
24006                     0.0          0.0            0.000000   
24045                     0.0          0.0            0.000000   
24404                     0.0          0.0            0.000000   
24452                     0.0          0.0            0.000000   
24454                     0.0          0.0            0.000000   
24598                     0.0          0.0            0.000000   
24608                     0.0          0.0            0.000000   
29324                     0.0          0.0            2.000000   

       Fwd Packet Length Max  Subflow Fwd Packets  Flow Packets/s  \
2130                       0                    2             inf   
6683                       6                    1             inf   
9265                      14                    2             inf   
10138                      6                    2             inf   
11793                      6                    2             inf   
14833                      6                    1             inf   
15627                      6                    1             inf   
16361                      0                    2             inf   
20043                      0                    2             inf   
20524                      0                    2             inf   
20650                      0                    2             inf   
20779                      0                    2             inf   
20839                      0                    2             inf   
21003                      0                    2             inf   
21449                      0                    2             inf   
21700                      0                    2             inf   
22152                      0                    2             inf   
22955                      0                    2             inf   
23687                      0                    2             inf   
24006                      0                    2             inf   
24045                      0                    2             inf   
24404                      0                    2             inf   
24452                      0                    2             inf   
24454                      0                    2             inf   
24598                      0                    2             inf   
24608                      0                    2             inf   
29324                      0                    1             inf   

       Total Fwd Packets  Destination Port  Label  
2130                   2             52606      0  
6683                   1             64231      0  
9265                   2             47542      0  
10138                  2             17774      0  
11793                  2                21      0  
14833                  1             64299      0  
15627                  1               443      0  
16361                  2             19152      0  
20043                  2                80      2  
20524                  2                80      2  
20650                  2                80      2  
20779                  2                80      2  
20839                  2                80      2  
21003                  2                80      2  
21449                  2                80      2  
21700                  2                80      2  
22152                  2                80      2  
22955                  2                80      2  
23687                  2                80      2  
24006                  2                80      2  
24045                  2                80      2  
24404                  2                80      2  
24452                  2                80      2  
24454                  2                80      2  
24598                  2                80      2  
24608                  2                80      2  
29324                  1              1026      3  

Features with infinite values:
['Flow Bytes/s', 'Flow Packets/s']

# Since there are not many rows that have inf as value we can simply decide to drop them (even if they are form a class that is not really present in the dataset)
df.replace([np.inf, -np.inf], np.nan, inplace=True)
df.dropna(inplace=True)

import sklearn
from sklearn.preprocessing import LabelEncoder
label_encoder = LabelEncoder()
df['Label'] = label_encoder.fit_transform(df['Label'])
#df['Destination Port'] = label_encoder.fit_transform(df['Destination Port'])
label_mapping = dict(zip(label_encoder.classes_, label_encoder.transform(label_encoder.classes_)))


print(label_mapping)

df.head()

{np.int64(0): np.int64(0), np.int64(1): np.int64(1), np.int64(2): np.int64(2), np.int64(3): np.int64(3)}

df[df['Flow Packets/s'] <0]

import pandas as pd



rows_with_negative_values = df[(df < 0).any(axis=1)]


num_rows_with_negative_values = len(rows_with_negative_values)


print(f"Number of rows with negative values: {num_rows_with_negative_values}")


print("\nRows with negative values:")
print(rows_with_negative_values)

Number of rows with negative values: 2

Rows with negative values:
       Flow Duration  Flow IAT Mean  Fwd PSH Flags  Bwd Packet Length Mean  \
12379             -1           -1.0              0                     6.0   
15783             -1           -1.0              0                     6.0   

       Bwd Packet Length Max  Flow Bytes/s  Down/Up Ratio  SYN Flag Count  \
12379                      6   -12000000.0              1               0   
15783                      6   -12000000.0              1               0   

       Fwd Packet Length Mean  Fwd IAT Std  Packet Length Mean  \
12379                     6.0          0.0                 6.0   
15783                     6.0          0.0                 6.0   

       Fwd Packet Length Max  Subflow Fwd Packets  Flow Packets/s  \
12379                      6                    1      -2000000.0   
15783                      6                    1      -2000000.0   

       Total Fwd Packets  Destination Port  Label  
12379                  1               443      0  
15783                  1                80      0

# Drop rows with at least one negative value in any column
df = df[(df >= 0).all(axis=1)]


df = df.reset_index(drop=True)

duplicate_rows = df[df.duplicated()]

print("Duplicate Rows:")
duplicate_rows.groupby('Label').count()
#2094 rows (majority DoS Hulk )

Duplicate Rows:

df.drop_duplicates(inplace=True)

# From here all the columns have decents stats, the only thing that could be done is to look for outliers in certain colums

df.describe()

# other thing to take into consideration is that we did not change the port into a numerical feature
#but like we said it does not matter the port is a bias for certain classes

#{'Benign': np.int64(0), 'Brute Force': np.int64(1), 'DoS Hulk': np.int64(2), 'PortScan': np.int64(3)}

filtered_df = df[df['Label'] == 3]
filtered_df

# unbalanced dataset
label_counts = df['Label'].value_counts()
print(label_counts)

Label
0    19240
3     4849
2     3868
1     1427
Name: count, dtype: int64

i = 1
# plot each column
plt.figure(figsize=(25,20))
for c in df.columns:
    plt.subplot(len(df.columns), 1, i)
    plt.plot(df[c])
    plt.title(c, y=0.75, loc='right', fontsize = 18)
    i += 1
plt.show()
plt.close()

import matplotlib.pyplot as plt
import seaborn as sns



sns.set_style("whitegrid")

# 1. Scatter Plot Matrix (Pairplot)
features_to_plot = ['Flow Duration', 'Flow Bytes/s', 'Flow Packets/s', 'Total Fwd Packets', 'Bwd Packet Length Mean']
plt.figure(figsize=(12, 8))
sns.pairplot(df[features_to_plot + ['Label']], hue='Label', diag_kind='kde', palette='coolwarm', plot_kws={'alpha': 0.6})
plt.suptitle("Pairplot of Selected Features", fontsize=16, y=1.02)
plt.show()


features_x = ['Flow Duration', 'Flow Bytes/s', 'Flow Packets/s', 'Fwd Packet Length Mean']
features_y = ['Total Fwd Packets', 'Bwd Packet Length Mean', 'Packet Length Mean', 'Fwd IAT Std']

plt.figure(figsize=(18, 16))
plot_index = 1
for feature1 in features_x:
    for feature2 in features_y:
        plt.subplot(len(features_x), len(features_y), plot_index)
        sns.scatterplot(x=df[feature1], y=df[feature2], hue=df['Label'], alpha=0.6, palette='coolwarm', edgecolor=None)
        plt.xlabel(feature1, fontsize=10)
        plt.ylabel(feature2, fontsize=10)
        plt.title(f'{feature1} vs {feature2}', fontsize=12)
        plot_index += 1

plt.suptitle("Scatter Plots of Selected Features", fontsize=16, y=1.02)
plt.tight_layout()
plt.show()

<Figure size 1200x800 with 0 Axes>

import matplotlib.pyplot as plt
import seaborn as sns

# Correlation Matrix
corr_matrix = df.corr()


plt.figure(figsize=(12, 10))
sns.heatmap(corr_matrix, annot=True, cmap='coolwarm', fmt=".2f", square=True)
plt.title("Correlation Matrix Heatmap")
plt.show()

import numpy as np


THRESHOLD = 0.99

corr = df.corr().abs()
upper = corr.where(np.triu(np.ones(corr.shape), k=1).astype(bool))

to_drop = [col for col in upper.columns if any(upper[col] > THRESHOLD)]
print(to_drop)

['SYN Flag Count', 'Total Fwd Packets']

(df['Fwd PSH Flags'] == df['SYN Flag Count']).all()

np.True_

df[['Fwd PSH Flags', 'SYN Flag Count']].drop_duplicates()

cols_to_drop = [
    'Fwd PSH Flags',
    'Subflow Fwd Packets'
]

df = df.drop(columns=cols_to_drop)

import matplotlib.pyplot as plt
import seaborn as sns

# Number of occurrences for each class
class_counts = df['Label'].value_counts().sort_index()
print("Class counts:")
print(class_counts)


plt.figure(figsize=(8, 6))
sns.countplot(x='Label', data=df, legend=False)
plt.title("Class Distribution")
plt.xlabel("Label (Encoded)")
plt.ylabel("Count")
plt.show()
#{'Benign': np.int64(0), 'Brute Force': np.int64(1), 'DoS Hulk': np.int64(2), 'PortScan': np.int64(3)}

Class counts:
Label
0    19240
1     1427
2     3868
3     4849
Name: count, dtype: int64

import matplotlib.pyplot as plt
import seaborn as sns


features = ['Flow Duration', 'Flow Bytes/s', 'Flow Packets/s', 'Bwd Packet Length Mean', 'Packet Length Mean', 'Fwd IAT Std']


plt.figure(figsize=(15, 10))
for i, feature in enumerate(features, 1):
    plt.subplot(4, 4, i)
    sns.boxplot(x='Label', y=feature, data=df, palette='Set2')
    plt.title(f'Boxplot di {feature} per Class')
    plt.xlabel('Class')
    plt.ylabel(feature)

plt.tight_layout()
plt.show()

/tmp/ipython-input-39-4198350486.py:11: FutureWarning: 

Passing `palette` without assigning `hue` is deprecated and will be removed in v0.14.0. Assign the `x` variable to `hue` and set `legend=False` for the same effect.

  sns.boxplot(x='Label', y=feature, data=df, palette='Set2')
/tmp/ipython-input-39-4198350486.py:11: FutureWarning: 

Passing `palette` without assigning `hue` is deprecated and will be removed in v0.14.0. Assign the `x` variable to `hue` and set `legend=False` for the same effect.

  sns.boxplot(x='Label', y=feature, data=df, palette='Set2')
/tmp/ipython-input-39-4198350486.py:11: FutureWarning: 

Passing `palette` without assigning `hue` is deprecated and will be removed in v0.14.0. Assign the `x` variable to `hue` and set `legend=False` for the same effect.

  sns.boxplot(x='Label', y=feature, data=df, palette='Set2')
/tmp/ipython-input-39-4198350486.py:11: FutureWarning: 

Passing `palette` without assigning `hue` is deprecated and will be removed in v0.14.0. Assign the `x` variable to `hue` and set `legend=False` for the same effect.

  sns.boxplot(x='Label', y=feature, data=df, palette='Set2')
/tmp/ipython-input-39-4198350486.py:11: FutureWarning: 

Passing `palette` without assigning `hue` is deprecated and will be removed in v0.14.0. Assign the `x` variable to `hue` and set `legend=False` for the same effect.

  sns.boxplot(x='Label', y=feature, data=df, palette='Set2')
/tmp/ipython-input-39-4198350486.py:11: FutureWarning: 

Passing `palette` without assigning `hue` is deprecated and will be removed in v0.14.0. Assign the `x` variable to `hue` and set `legend=False` for the same effect.

  sns.boxplot(x='Label', y=feature, data=df, palette='Set2')

plt.figure(figsize=(15, 10))
for i, feature in enumerate(features, 1):
    plt.subplot(4, 4, i)
    sns.violinplot(x='Label', y=feature, data=df, palette='Set2', inner='quartile')
    plt.title(f'Violin plot di {feature} per Classe')
    plt.xlabel('Classe')
    plt.ylabel(feature)

plt.tight_layout()
plt.show()

/tmp/ipython-input-40-3136383551.py:4: FutureWarning: 

Passing `palette` without assigning `hue` is deprecated and will be removed in v0.14.0. Assign the `x` variable to `hue` and set `legend=False` for the same effect.

  sns.violinplot(x='Label', y=feature, data=df, palette='Set2', inner='quartile')
/tmp/ipython-input-40-3136383551.py:4: FutureWarning: 

Passing `palette` without assigning `hue` is deprecated and will be removed in v0.14.0. Assign the `x` variable to `hue` and set `legend=False` for the same effect.

  sns.violinplot(x='Label', y=feature, data=df, palette='Set2', inner='quartile')
/tmp/ipython-input-40-3136383551.py:4: FutureWarning: 

Passing `palette` without assigning `hue` is deprecated and will be removed in v0.14.0. Assign the `x` variable to `hue` and set `legend=False` for the same effect.

  sns.violinplot(x='Label', y=feature, data=df, palette='Set2', inner='quartile')
/tmp/ipython-input-40-3136383551.py:4: FutureWarning: 

Passing `palette` without assigning `hue` is deprecated and will be removed in v0.14.0. Assign the `x` variable to `hue` and set `legend=False` for the same effect.

  sns.violinplot(x='Label', y=feature, data=df, palette='Set2', inner='quartile')
/tmp/ipython-input-40-3136383551.py:4: FutureWarning: 

Passing `palette` without assigning `hue` is deprecated and will be removed in v0.14.0. Assign the `x` variable to `hue` and set `legend=False` for the same effect.

  sns.violinplot(x='Label', y=feature, data=df, palette='Set2', inner='quartile')
/tmp/ipython-input-40-3136383551.py:4: FutureWarning: 

Passing `palette` without assigning `hue` is deprecated and will be removed in v0.14.0. Assign the `x` variable to `hue` and set `legend=False` for the same effect.

  sns.violinplot(x='Label', y=feature, data=df, palette='Set2', inner='quartile')

numeric_feats = df.select_dtypes(include=['int64','float64']).columns
skewness = df[numeric_feats].skew().sort_values(ascending=False)
print(skewness)

Total Fwd Packets         152.007747
Flow Bytes/s               41.798570
Fwd Packet Length Max      12.173963
Fwd Packet Length Mean     11.139485
Flow Packets/s              7.280535
Flow IAT Mean               7.197015
SYN Flag Count              4.737087
Fwd IAT Std                 2.526223
Bwd Packet Length Max       2.465285
Destination Port            2.154604
Bwd Packet Length Mean      2.105616
Packet Length Mean          1.980722
Flow Duration               1.810170
Label                       0.968786
Down/Up Ratio               0.347508
dtype: float64

import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import FunctionTransformer, RobustScaler
import matplotlib.pyplot as plt
import seaborn as sns


X = df.drop(columns=['Label'])
y = df['Label']

# 1.  training (60%), validation (20%), and test (20%) sets
X_train, X_temp, y_train, y_temp = train_test_split(
    X, y, test_size=0.4, random_state=42, stratify=y
)
X_val, X_test, y_val, y_test = train_test_split(
    X_temp, y_temp, test_size=0.5, random_state=42, stratify=y_temp
)


print("Class distribution - Train:\n", y_train.value_counts())
print("Class distribution - Val:\n", y_val.value_counts())
print("Class distribution - Test:\n", y_test.value_counts())

Class distribution - Train:
 Label
0    11544
3     2909
2     2321
1      856
Name: count, dtype: int64
Class distribution - Val:
 Label
0    3848
3     970
2     774
1     285
Name: count, dtype: int64
Class distribution - Test:
 Label
0    3848
3     970
2     773
1     286
Name: count, dtype: int64

import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
from scipy import stats
import scipy.stats as stats_scipy

def analyze_class_distribution(y_train, y_test, y_val):
    """
    Analyze class distribution across train, test, and validation sets.

    Args:
        y_train (np.ndarray): Training labels.
        y_test (np.ndarray): Test labels.
        y_val (np.ndarray): Validation labels.
    """
    def get_class_distribution(labels):
        unique, counts = np.unique(labels, return_counts=True)
        total = len(labels)
        percentages = (counts / total) * 100
        return dict(zip(unique, percentages))

    print("Train Set Class Distribution:")
    train_dist = get_class_distribution(y_train)
    for cls, percentage in train_dist.items():
        print(f"Class {cls}: {percentage:.2f}%")

    print("\nTest Set Class Distribution:")
    test_dist = get_class_distribution(y_test)
    for cls, percentage in test_dist.items():
        print(f"Class {cls}: {percentage:.2f}%")

    print("\nValidation Set Class Distribution:")
    val_dist = get_class_distribution(y_val)
    for cls, percentage in val_dist.items():
        print(f"Class {cls}: {percentage:.2f}%")

    return train_dist, test_dist, val_dist

distributions = analyze_class_distribution(y_train, y_test, y_val)
print(distributions)

def compare_distributions(train_data, test_data, val_data, feature_name):
    """
    Compare distributions of a specific feature across train, test, and validation sets.

    Parameters:
        train_data: Training dataset feature column.
        test_data: Test dataset feature column.
        val_data: Validation dataset feature column.
        feature_name: Name of the feature being analyzed.

    Returns:
        Dictionary of statistical test results and visualization.
    """
    results = {}

    # 1. Statistical Tests (Kolmogorov-Smirnov Test)
    def ks_test_between_sets(set1, set2, set_names):
        statistic, p_value = stats.ks_2samp(set1, set2)
        return {
            'statistic': statistic,
            'p_value': p_value,
            'sets_compared': set_names,
            'significant_difference': p_value < 0.05
        }

    results['ks_tests'] = {
        'train_vs_test': ks_test_between_sets(train_data, test_data, 'Train vs Test'),
        'train_vs_val': ks_test_between_sets(train_data, val_data, 'Train vs Validation'),
        'test_vs_val': ks_test_between_sets(test_data, val_data, 'Test vs Validation')
    }

    # 2. Distribution Visualization
    plt.figure(figsize=(15, 5))
    plt.subplot(131)
    plt.hist(train_data, bins=30, alpha=0.5, label='Train')
    plt.hist(test_data, bins=30, alpha=0.5, label='Test')
    plt.hist(val_data, bins=30, alpha=0.5, label='Validation')
    plt.title(f'{feature_name} Distribution')
    plt.xlabel(feature_name)
    plt.ylabel('Frequency')
    plt.legend()

    plt.subplot(132)
    box_data = [train_data, test_data, val_data]
    plt.boxplot(box_data, labels=['Train', 'Test', 'Validation'])
    plt.title(f'{feature_name} Box Plot')
    plt.ylabel(feature_name)

    plt.subplot(133)
    sns.kdeplot(train_data, label='Train')
    sns.kdeplot(test_data, label='Test')
    sns.kdeplot(val_data, label='Validation')
    plt.title(f'{feature_name} Density Plot')
    plt.xlabel(feature_name)
    plt.ylabel('Density')
    plt.legend()
    plt.tight_layout()
    plt.show()

    # 3. Distribution Metrics (Jensen-Shannon Divergence)
    def jensen_shannon_divergence(p, q):
        p = np.asarray(p)
        q = np.asarray(q)
        m = 0.5 * (p + q)
        return 0.5 * (stats_scipy.entropy(p, m) + stats_scipy.entropy(q, m))

    def get_distribution_histogram(data, bins=30):
        hist, _ = np.histogram(data, bins=bins, density=True)
        return hist

    results['distribution_metrics'] = {
        'train_vs_test_js_div': jensen_shannon_divergence(
            get_distribution_histogram(train_data),
            get_distribution_histogram(test_data)
        ),
        'train_vs_val_js_div': jensen_shannon_divergence(
            get_distribution_histogram(train_data),
            get_distribution_histogram(val_data)
        ),
        'test_vs_val_js_div': jensen_shannon_divergence(
            get_distribution_histogram(test_data),
            get_distribution_histogram(val_data)
        )
    }
    return results

def analyze_feature_distributions(X_train, X_test, X_val, feature_names=None):
    """
    Analyze distributions for multiple features.

    Parameters:
        X_train: Training dataset (2D NumPy array).
        X_test: Test dataset (2D NumPy array).
        X_val: Validation dataset (2D NumPy array).
        feature_names: Optional list of feature names (default is column indices).
    """
    if feature_names is None:
        feature_names = [f'Feature_{i}' for i in range(X_train.shape[1])]

    assert X_train.shape[1] == len(feature_names), "Number of feature names must match number of columns"
    comprehensive_results = {}
    for idx, feature in enumerate(feature_names):
        print(f"\n--- Analysis for Feature: {feature} ---")
        train_feature = X_train[:, idx]
        test_feature = X_test[:, idx]
        val_feature = X_val[:, idx]
        results = compare_distributions(train_feature, test_feature, val_feature, feature)
        comprehensive_results[feature] = results
        print("Kolmogorov-Smirnov Test Results:")
        for key, test_result in results['ks_tests'].items():
            print(f"{test_result['sets_compared']}:")
            print(f"  Statistic: {test_result['statistic']:.4f}")
            print(f"  P-value: {test_result['p_value']:.4f}")
            print(f"  Significant Difference: {test_result['significant_difference']}")
        print("\nJensen-Shannon Divergence:")
        for key, div_value in results['distribution_metrics'].items():
            print(f"{key}: {div_value:.4f}")
    return comprehensive_results

features = X_train.columns.tolist()
analyze_feature_distributions(X_train.to_numpy(), X_test.to_numpy(), X_val.to_numpy(), features)

Train Set Class Distribution:
Class 0: 65.48%
Class 1: 4.86%
Class 2: 13.17%
Class 3: 16.50%

Test Set Class Distribution:
Class 0: 65.48%
Class 1: 4.87%
Class 2: 13.15%
Class 3: 16.51%

Validation Set Class Distribution:
Class 0: 65.48%
Class 1: 4.85%
Class 2: 13.17%
Class 3: 16.51%
({np.int64(0): np.float64(65.47929665343165), np.int64(1): np.float64(4.855360181508791), np.int64(2): np.float64(13.16505955757232), np.int64(3): np.float64(16.500283607487237)}, {np.int64(0): np.float64(65.4755827803301), np.int64(1): np.float64(4.8664284498893995), np.int64(2): np.float64(13.152969201973796), np.int64(3): np.float64(16.505019567806702)}, {np.int64(0): np.float64(65.4755827803301), np.int64(1): np.float64(4.849412965798877), np.int64(2): np.float64(13.169984686064318), np.int64(3): np.float64(16.505019567806702)})

--- Analysis for Feature: Flow Duration ---

/tmp/ipython-input-363-3966410394.py:86: MatplotlibDeprecationWarning: The 'labels' parameter of boxplot() has been renamed 'tick_labels' since Matplotlib 3.9; support for the old name will be dropped in 3.11.
  plt.boxplot(box_data, labels=['Train', 'Test', 'Validation'])

Kolmogorov-Smirnov Test Results:
Train vs Test:
  Statistic: 0.0075
  P-value: 0.9632
  Significant Difference: False
Train vs Validation:
  Statistic: 0.0136
  P-value: 0.3831
  Significant Difference: False
Test vs Validation:
  Statistic: 0.0133
  P-value: 0.6787
  Significant Difference: False

Jensen-Shannon Divergence:
train_vs_test_js_div: 0.0010
train_vs_val_js_div: 0.0008
test_vs_val_js_div: 0.0008

--- Analysis for Feature: Flow IAT Mean ---

/tmp/ipython-input-363-3966410394.py:86: MatplotlibDeprecationWarning: The 'labels' parameter of boxplot() has been renamed 'tick_labels' since Matplotlib 3.9; support for the old name will be dropped in 3.11.
  plt.boxplot(box_data, labels=['Train', 'Test', 'Validation'])

Kolmogorov-Smirnov Test Results:
Train vs Test:
  Statistic: 0.0108
  P-value: 0.6738
  Significant Difference: False
Train vs Validation:
  Statistic: 0.0121
  P-value: 0.5356
  Significant Difference: False
Test vs Validation:
  Statistic: 0.0131
  P-value: 0.6942
  Significant Difference: False

Jensen-Shannon Divergence:
train_vs_test_js_div: 0.0370
train_vs_val_js_div: 0.0391
test_vs_val_js_div: 0.0008

--- Analysis for Feature: Bwd Packet Length Mean ---

/tmp/ipython-input-363-3966410394.py:86: MatplotlibDeprecationWarning: The 'labels' parameter of boxplot() has been renamed 'tick_labels' since Matplotlib 3.9; support for the old name will be dropped in 3.11.
  plt.boxplot(box_data, labels=['Train', 'Test', 'Validation'])

Kolmogorov-Smirnov Test Results:
Train vs Test:
  Statistic: 0.0085
  P-value: 0.9050
  Significant Difference: False
Train vs Validation:
  Statistic: 0.0142
  P-value: 0.3313
  Significant Difference: False
Test vs Validation:
  Statistic: 0.0143
  P-value: 0.5857
  Significant Difference: False

Jensen-Shannon Divergence:
train_vs_test_js_div: 0.0774
train_vs_val_js_div: 0.0780
test_vs_val_js_div: 0.0010

--- Analysis for Feature: Bwd Packet Length Max ---

/tmp/ipython-input-363-3966410394.py:86: MatplotlibDeprecationWarning: The 'labels' parameter of boxplot() has been renamed 'tick_labels' since Matplotlib 3.9; support for the old name will be dropped in 3.11.
  plt.boxplot(box_data, labels=['Train', 'Test', 'Validation'])

Kolmogorov-Smirnov Test Results:
Train vs Test:
  Statistic: 0.0090
  P-value: 0.8634
  Significant Difference: False
Train vs Validation:
  Statistic: 0.0149
  P-value: 0.2757
  Significant Difference: False
Test vs Validation:
  Statistic: 0.0160
  P-value: 0.4399
  Significant Difference: False

Jensen-Shannon Divergence:
train_vs_test_js_div: 0.0005
train_vs_val_js_div: 0.0013
test_vs_val_js_div: 0.0015

--- Analysis for Feature: Flow Bytes/s ---

/tmp/ipython-input-363-3966410394.py:86: MatplotlibDeprecationWarning: The 'labels' parameter of boxplot() has been renamed 'tick_labels' since Matplotlib 3.9; support for the old name will be dropped in 3.11.
  plt.boxplot(box_data, labels=['Train', 'Test', 'Validation'])

import numpy as np

def detect_outliers_iqr(data, multiplier=1.5):
    """
    Detect outliers using Interquartile Range method

    Parameters:
    - data: NumPy array
    - multiplier: IQR multiplier (default 1.5)

    Returns:
    - Boolean array indicating outliers
    """
    Q1 = np.percentile(data, 25)
    Q3 = np.percentile(data, 75)
    IQR = Q3 - Q1
    lower_bound = Q1 - (multiplier * IQR)
    upper_bound = Q3 + (multiplier * IQR)
    return (data < lower_bound) | (data > upper_bound)

def analyze_outliers_by_class(X, y, method='iqr', multiplier=1.5):
    """
    Analyze outliers across different classes

    Parameters:
    - X: Feature matrix (2D NumPy array)
    - y: Class labels (1D NumPy array)
    - method: Outlier detection method ('iqr')
    - multiplier: Multiplier for IQR method

    Returns:
    - Comprehensive outlier analysis by class
    """
    # Unique classes
    unique_classes = np.unique(y)

    # Outlier analysis results
    outlier_analysis = {
        'overall_results': {},
        'class_results': {}
    }

    # Total outliers tracking
    total_outliers = 0
    total_samples = len(y)

    # Analyze each feature
    for feature_idx in range(X.shape[1]):
        feature_outliers = {
            'overall_outliers': [],
            'class_outliers': {}
        }

        # Current feature column
        feature_data = X[:, feature_idx]

        # Detect overall outliers
        overall_outlier_mask = detect_outliers_iqr(feature_data, multiplier)
        feature_outliers['overall_outliers'] = overall_outlier_mask

        # Per-class analysis
        for cls in unique_classes:
            # Mask for current class
            class_mask = (y == cls)

            # Feature data for current class
            class_feature_data = feature_data[class_mask]

            # Detect outliers for this class
            class_outlier_mask = detect_outliers_iqr(class_feature_data, multiplier)

            # Store class-specific results
            feature_outliers['class_outliers'][cls] = {
                'total_samples': len(class_feature_data),
                'outlier_samples': np.sum(class_outlier_mask),
                'outlier_percentage': (np.sum(class_outlier_mask) / len(class_feature_data)) * 100
            }

        # Store results for this feature
        outlier_analysis['overall_results'][f'Feature_{feature_idx}'] = feature_outliers

    # Summarize overall results
    print("\nOutlier Analysis Summary:")
    for feature, results in outlier_analysis['overall_results'].items():
        print(f"\n{feature}:")

        # Overall outliers
        overall_outliers = np.sum(results['overall_outliers'])
        overall_outlier_percentage = (overall_outliers / total_samples) * 100
        print(f"  Total Outliers: {overall_outliers} ({overall_outlier_percentage:.2f}%)")

        # Class-specific results
        print("  Class-wise Outlier Breakdown:")
        for cls, class_results in results['class_outliers'].items():
            print(f"    Class {cls}:")
            print(f"      Total Samples: {class_results['total_samples']}")
            print(f"      Outlier Samples: {class_results['outlier_samples']}")
            print(f"      Outlier Percentage: {class_results['outlier_percentage']:.2f}%")

    return outlier_analysis





outlier_analysis = analyze_outliers_by_class(X_train.to_numpy(), y_train)

Outlier Analysis Summary:

Feature_0:
  Total Outliers: 3483 (19.76%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 2669
      Outlier Percentage: 23.12%
    Class 1:
      Total Samples: 856
      Outlier Samples: 128
      Outlier Percentage: 14.95%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 485
      Outlier Percentage: 20.90%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 348
      Outlier Percentage: 11.96%

Feature_1:
  Total Outliers: 2898 (16.44%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 2512
      Outlier Percentage: 21.76%
    Class 1:
      Total Samples: 856
      Outlier Samples: 115
      Outlier Percentage: 13.43%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 499
      Outlier Percentage: 21.50%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 348
      Outlier Percentage: 11.96%

Feature_2:
  Total Outliers: 3183 (18.05%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 1376
      Outlier Percentage: 11.92%
    Class 1:
      Total Samples: 856
      Outlier Samples: 93
      Outlier Percentage: 10.86%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 368
      Outlier Percentage: 15.86%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 16
      Outlier Percentage: 0.55%

Feature_3:
  Total Outliers: 3901 (22.13%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 2306
      Outlier Percentage: 19.98%
    Class 1:
      Total Samples: 856
      Outlier Samples: 93
      Outlier Percentage: 10.86%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 577
      Outlier Percentage: 24.86%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 16
      Outlier Percentage: 0.55%

Feature_4:
  Total Outliers: 2846 (16.14%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 2386
      Outlier Percentage: 20.67%
    Class 1:
      Total Samples: 856
      Outlier Samples: 93
      Outlier Percentage: 10.86%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 529
      Outlier Percentage: 22.79%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 349
      Outlier Percentage: 12.00%

Feature_5:
  Total Outliers: 49 (0.28%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 49
      Outlier Percentage: 0.42%
    Class 1:
      Total Samples: 856
      Outlier Samples: 81
      Outlier Percentage: 9.46%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 0
      Outlier Percentage: 0.00%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 15
      Outlier Percentage: 0.52%

Feature_6:
  Total Outliers: 714 (4.05%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 714
      Outlier Percentage: 6.19%
    Class 1:
      Total Samples: 856
      Outlier Samples: 0
      Outlier Percentage: 0.00%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 0
      Outlier Percentage: 0.00%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 0
      Outlier Percentage: 0.00%

Feature_7:
  Total Outliers: 998 (5.66%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 939
      Outlier Percentage: 8.13%
    Class 1:
      Total Samples: 856
      Outlier Samples: 93
      Outlier Percentage: 10.86%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 203
      Outlier Percentage: 8.75%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 1
      Outlier Percentage: 0.03%

Feature_8:
  Total Outliers: 2415 (13.70%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 2520
      Outlier Percentage: 21.83%
    Class 1:
      Total Samples: 856
      Outlier Samples: 133
      Outlier Percentage: 15.54%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 484
      Outlier Percentage: 20.85%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 8
      Outlier Percentage: 0.28%

Feature_9:
  Total Outliers: 2951 (16.74%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 1450
      Outlier Percentage: 12.56%
    Class 1:
      Total Samples: 856
      Outlier Samples: 93
      Outlier Percentage: 10.86%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 450
      Outlier Percentage: 19.39%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 9
      Outlier Percentage: 0.31%

Feature_10:
  Total Outliers: 845 (4.79%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 2712
      Outlier Percentage: 23.49%
    Class 1:
      Total Samples: 856
      Outlier Samples: 93
      Outlier Percentage: 10.86%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 319
      Outlier Percentage: 13.74%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 8
      Outlier Percentage: 0.28%

Feature_11:
  Total Outliers: 1416 (8.03%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 1079
      Outlier Percentage: 9.35%
    Class 1:
      Total Samples: 856
      Outlier Samples: 104
      Outlier Percentage: 12.15%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 496
      Outlier Percentage: 21.37%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 488
      Outlier Percentage: 16.78%

Feature_12:
  Total Outliers: 1498 (8.50%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 2191
      Outlier Percentage: 18.98%
    Class 1:
      Total Samples: 856
      Outlier Samples: 129
      Outlier Percentage: 15.07%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 1
      Outlier Percentage: 0.04%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 23
      Outlier Percentage: 0.79%

Feature_13:
  Total Outliers: 3902 (22.13%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 2322
      Outlier Percentage: 20.11%
    Class 1:
      Total Samples: 856
      Outlier Samples: 0
      Outlier Percentage: 0.00%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 0
      Outlier Percentage: 0.00%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 364
      Outlier Percentage: 12.51%

import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
import matplotlib.pyplot as plt
import seaborn as sns



print("Class distribution - Train:\n", y_train.value_counts())
print("Class distribution - Validation:\n", y_val.value_counts())
print("Class distribution - Test:\n", y_test.value_counts())


scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train)
X_val_scaled = scaler.transform(X_val)


X_train_scaled = pd.DataFrame(X_train_scaled, columns=X_train.columns, index=X_train.index)
X_val_scaled = pd.DataFrame(X_val_scaled, columns=X_val.columns, index=X_val.index)


X_test_scaled= scaler.transform(X_test)
X_test_scaled = pd.DataFrame(X_test_scaled, columns=X_test.columns, index=X_test.index)

Class distribution - Train:
 Label
0    11544
3     2909
2     2321
1      856
Name: count, dtype: int64
Class distribution - Validation:
 Label
0    3848
3     970
2     774
1     285
Name: count, dtype: int64
Class distribution - Test:
 Label
0    3848
3     970
2     773
1     286
Name: count, dtype: int64

import numpy as np

def detect_outliers_iqr(data, multiplier=1.5):
    """
    Detect outliers using Interquartile Range method

    Parameters:
    - data: NumPy array
    - multiplier: IQR multiplier (default 1.5)

    Returns:
    - Boolean array indicating outliers
    """
    Q1 = np.percentile(data, 25)
    Q3 = np.percentile(data, 75)
    IQR = Q3 - Q1
    lower_bound = Q1 - (multiplier * IQR)
    upper_bound = Q3 + (multiplier * IQR)
    return (data < lower_bound) | (data > upper_bound)

def analyze_outliers_by_class(X, y, method='iqr', multiplier=1.5):
    """
    Analyze outliers across different classes

    Parameters:
    - X: Feature matrix (2D NumPy array)
    - y: Class labels (1D NumPy array)
    - method: Outlier detection method ('iqr')
    - multiplier: Multiplier for IQR method

    Returns:
    - Comprehensive outlier analysis by class
    """
    # Unique classes
    unique_classes = np.unique(y)

    # Outlier analysis results
    outlier_analysis = {
        'overall_results': {},
        'class_results': {}
    }

    # Total outliers tracking
    total_outliers = 0
    total_samples = len(y)

    # Analyze each feature
    for feature_idx in range(X.shape[1]):
        feature_outliers = {
            'overall_outliers': [],
            'class_outliers': {}
        }

        # Current feature column
        feature_data = X[:, feature_idx]

        # Detect overall outliers
        overall_outlier_mask = detect_outliers_iqr(feature_data, multiplier)
        feature_outliers['overall_outliers'] = overall_outlier_mask

        # Per-class analysis
        for cls in unique_classes:
            # Mask for current class
            class_mask = (y == cls)

            # Feature data for current class
            class_feature_data = feature_data[class_mask]

            # Detect outliers for this class
            class_outlier_mask = detect_outliers_iqr(class_feature_data, multiplier)

            # Store class-specific results
            feature_outliers['class_outliers'][cls] = {
                'total_samples': len(class_feature_data),
                'outlier_samples': np.sum(class_outlier_mask),
                'outlier_percentage': (np.sum(class_outlier_mask) / len(class_feature_data)) * 100
            }

        # Store results for this feature
        outlier_analysis['overall_results'][f'Feature_{feature_idx}'] = feature_outliers

    # Summarize overall results
    print("\nOutlier Analysis Summary:")
    for feature, results in outlier_analysis['overall_results'].items():
        print(f"\n{feature}:")

        # Overall outliers
        overall_outliers = np.sum(results['overall_outliers'])
        overall_outlier_percentage = (overall_outliers / total_samples) * 100
        print(f"  Total Outliers: {overall_outliers} ({overall_outlier_percentage:.2f}%)")

        # Class-specific results
        print("  Class-wise Outlier Breakdown:")
        for cls, class_results in results['class_outliers'].items():
            print(f"    Class {cls}:")
            print(f"      Total Samples: {class_results['total_samples']}")
            print(f"      Outlier Samples: {class_results['outlier_samples']}")
            print(f"      Outlier Percentage: {class_results['outlier_percentage']:.2f}%")

    return outlier_analysis





outlier_analysis = analyze_outliers_by_class(X_train_scaled.to_numpy(), y_train)

Outlier Analysis Summary:

Feature_0:
  Total Outliers: 3483 (19.76%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 2669
      Outlier Percentage: 23.12%
    Class 1:
      Total Samples: 856
      Outlier Samples: 128
      Outlier Percentage: 14.95%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 485
      Outlier Percentage: 20.90%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 348
      Outlier Percentage: 11.96%

Feature_1:
  Total Outliers: 2898 (16.44%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 2512
      Outlier Percentage: 21.76%
    Class 1:
      Total Samples: 856
      Outlier Samples: 115
      Outlier Percentage: 13.43%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 499
      Outlier Percentage: 21.50%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 348
      Outlier Percentage: 11.96%

Feature_2:
  Total Outliers: 3183 (18.05%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 1376
      Outlier Percentage: 11.92%
    Class 1:
      Total Samples: 856
      Outlier Samples: 93
      Outlier Percentage: 10.86%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 368
      Outlier Percentage: 15.86%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 16
      Outlier Percentage: 0.55%

Feature_3:
  Total Outliers: 3901 (22.13%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 2306
      Outlier Percentage: 19.98%
    Class 1:
      Total Samples: 856
      Outlier Samples: 93
      Outlier Percentage: 10.86%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 577
      Outlier Percentage: 24.86%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 16
      Outlier Percentage: 0.55%

Feature_4:
  Total Outliers: 2846 (16.14%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 2386
      Outlier Percentage: 20.67%
    Class 1:
      Total Samples: 856
      Outlier Samples: 93
      Outlier Percentage: 10.86%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 529
      Outlier Percentage: 22.79%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 349
      Outlier Percentage: 12.00%

Feature_5:
  Total Outliers: 49 (0.28%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 49
      Outlier Percentage: 0.42%
    Class 1:
      Total Samples: 856
      Outlier Samples: 81
      Outlier Percentage: 9.46%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 0
      Outlier Percentage: 0.00%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 15
      Outlier Percentage: 0.52%

Feature_6:
  Total Outliers: 714 (4.05%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 714
      Outlier Percentage: 6.19%
    Class 1:
      Total Samples: 856
      Outlier Samples: 0
      Outlier Percentage: 0.00%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 0
      Outlier Percentage: 0.00%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 0
      Outlier Percentage: 0.00%

Feature_7:
  Total Outliers: 998 (5.66%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 939
      Outlier Percentage: 8.13%
    Class 1:
      Total Samples: 856
      Outlier Samples: 93
      Outlier Percentage: 10.86%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 203
      Outlier Percentage: 8.75%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 1
      Outlier Percentage: 0.03%

Feature_8:
  Total Outliers: 2415 (13.70%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 2520
      Outlier Percentage: 21.83%
    Class 1:
      Total Samples: 856
      Outlier Samples: 133
      Outlier Percentage: 15.54%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 484
      Outlier Percentage: 20.85%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 8
      Outlier Percentage: 0.28%

Feature_9:
  Total Outliers: 2951 (16.74%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 1450
      Outlier Percentage: 12.56%
    Class 1:
      Total Samples: 856
      Outlier Samples: 93
      Outlier Percentage: 10.86%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 450
      Outlier Percentage: 19.39%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 9
      Outlier Percentage: 0.31%

Feature_10:
  Total Outliers: 845 (4.79%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 2712
      Outlier Percentage: 23.49%
    Class 1:
      Total Samples: 856
      Outlier Samples: 93
      Outlier Percentage: 10.86%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 319
      Outlier Percentage: 13.74%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 8
      Outlier Percentage: 0.28%

Feature_11:
  Total Outliers: 1416 (8.03%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 1079
      Outlier Percentage: 9.35%
    Class 1:
      Total Samples: 856
      Outlier Samples: 104
      Outlier Percentage: 12.15%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 496
      Outlier Percentage: 21.37%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 488
      Outlier Percentage: 16.78%

Feature_12:
  Total Outliers: 1498 (8.50%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 2191
      Outlier Percentage: 18.98%
    Class 1:
      Total Samples: 856
      Outlier Samples: 129
      Outlier Percentage: 15.07%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 1
      Outlier Percentage: 0.04%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 23
      Outlier Percentage: 0.79%

Feature_13:
  Total Outliers: 3902 (22.13%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 11544
      Outlier Samples: 2322
      Outlier Percentage: 20.11%
    Class 1:
      Total Samples: 856
      Outlier Samples: 0
      Outlier Percentage: 0.00%
    Class 2:
      Total Samples: 2321
      Outlier Samples: 0
      Outlier Percentage: 0.00%
    Class 3:
      Total Samples: 2909
      Outlier Samples: 364
      Outlier Percentage: 12.51%

import numpy as np

def detect_outliers_iqr(data, multiplier=1.5):
    """
    Detect outliers using Interquartile Range method

    Parameters:
    - data: NumPy array
    - multiplier: IQR multiplier (default 1.5)

    Returns:
    - Boolean array indicating outliers
    """
    Q1 = np.percentile(data, 25)
    Q3 = np.percentile(data, 75)
    IQR = Q3 - Q1
    lower_bound = Q1 - (multiplier * IQR)
    upper_bound = Q3 + (multiplier * IQR)
    return (data < lower_bound) | (data > upper_bound)

def analyze_outliers_by_class(X, y, method='iqr', multiplier=1.5):
    """
    Analyze outliers across different classes

    Parameters:
    - X: Feature matrix (2D NumPy array)
    - y: Class labels (1D NumPy array)
    - method: Outlier detection method ('iqr')
    - multiplier: Multiplier for IQR method

    Returns:
    - Comprehensive outlier analysis by class
    """
    # Unique classes
    unique_classes = np.unique(y)

    # Outlier analysis results
    outlier_analysis = {
        'overall_results': {},
        'class_results': {}
    }

    # Total outliers tracking
    total_outliers = 0
    total_samples = len(y)

    # Analyze each feature
    for feature_idx in range(X.shape[1]):
        feature_outliers = {
            'overall_outliers': [],
            'class_outliers': {}
        }

        # Current feature column
        feature_data = X[:, feature_idx]

        # Detect overall outliers
        overall_outlier_mask = detect_outliers_iqr(feature_data, multiplier)
        feature_outliers['overall_outliers'] = overall_outlier_mask

        # Per-class analysis
        for cls in unique_classes:
            # Mask for current class
            class_mask = (y == cls)

            # Feature data for current class
            class_feature_data = feature_data[class_mask]

            # Detect outliers for this class
            class_outlier_mask = detect_outliers_iqr(class_feature_data, multiplier)

            # Store class-specific results
            feature_outliers['class_outliers'][cls] = {
                'total_samples': len(class_feature_data),
                'outlier_samples': np.sum(class_outlier_mask),
                'outlier_percentage': (np.sum(class_outlier_mask) / len(class_feature_data)) * 100
            }

        # Store results for this feature
        outlier_analysis['overall_results'][f'Feature_{feature_idx}'] = feature_outliers

    # Summarize overall results
    print("\nOutlier Analysis Summary:")
    for feature, results in outlier_analysis['overall_results'].items():
        print(f"\n{feature}:")

        # Overall outliers
        overall_outliers = np.sum(results['overall_outliers'])
        overall_outlier_percentage = (overall_outliers / total_samples) * 100
        print(f"  Total Outliers: {overall_outliers} ({overall_outlier_percentage:.2f}%)")

        # Class-specific results
        print("  Class-wise Outlier Breakdown:")
        for cls, class_results in results['class_outliers'].items():
            print(f"    Class {cls}:")
            print(f"      Total Samples: {class_results['total_samples']}")
            print(f"      Outlier Samples: {class_results['outlier_samples']}")
            print(f"      Outlier Percentage: {class_results['outlier_percentage']:.2f}%")

    return outlier_analysis



# Demonstration

outlier_analysis = analyze_outliers_by_class(X_val_scaled.to_numpy(), y_val)

# can be runned again

Outlier Analysis Summary:

Feature_0:
  Total Outliers: 1098 (18.68%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 3848
      Outlier Samples: 889
      Outlier Percentage: 23.10%
    Class 1:
      Total Samples: 285
      Outlier Samples: 39
      Outlier Percentage: 13.68%
    Class 2:
      Total Samples: 774
      Outlier Samples: 0
      Outlier Percentage: 0.00%
    Class 3:
      Total Samples: 970
      Outlier Samples: 107
      Outlier Percentage: 11.03%

Feature_1:
  Total Outliers: 1070 (18.21%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 3848
      Outlier Samples: 841
      Outlier Percentage: 21.86%
    Class 1:
      Total Samples: 285
      Outlier Samples: 34
      Outlier Percentage: 11.93%
    Class 2:
      Total Samples: 774
      Outlier Samples: 5
      Outlier Percentage: 0.65%
    Class 3:
      Total Samples: 970
      Outlier Samples: 107
      Outlier Percentage: 11.03%

Feature_2:
  Total Outliers: 1067 (18.16%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 3848
      Outlier Samples: 445
      Outlier Percentage: 11.56%
    Class 1:
      Total Samples: 285
      Outlier Samples: 26
      Outlier Percentage: 9.12%
    Class 2:
      Total Samples: 774
      Outlier Samples: 141
      Outlier Percentage: 18.22%
    Class 3:
      Total Samples: 970
      Outlier Samples: 9
      Outlier Percentage: 0.93%

Feature_3:
  Total Outliers: 1337 (22.75%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 3848
      Outlier Samples: 726
      Outlier Percentage: 18.87%
    Class 1:
      Total Samples: 285
      Outlier Samples: 26
      Outlier Percentage: 9.12%
    Class 2:
      Total Samples: 774
      Outlier Samples: 203
      Outlier Percentage: 26.23%
    Class 3:
      Total Samples: 970
      Outlier Samples: 9
      Outlier Percentage: 0.93%

Feature_4:
  Total Outliers: 935 (15.91%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 3848
      Outlier Samples: 787
      Outlier Percentage: 20.45%
    Class 1:
      Total Samples: 285
      Outlier Samples: 26
      Outlier Percentage: 9.12%
    Class 2:
      Total Samples: 774
      Outlier Samples: 215
      Outlier Percentage: 27.78%
    Class 3:
      Total Samples: 970
      Outlier Samples: 113
      Outlier Percentage: 11.65%

Feature_5:
  Total Outliers: 15 (0.26%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 3848
      Outlier Samples: 15
      Outlier Percentage: 0.39%
    Class 1:
      Total Samples: 285
      Outlier Samples: 25
      Outlier Percentage: 8.77%
    Class 2:
      Total Samples: 774
      Outlier Samples: 0
      Outlier Percentage: 0.00%
    Class 3:
      Total Samples: 970
      Outlier Samples: 9
      Outlier Percentage: 0.93%

Feature_6:
  Total Outliers: 232 (3.95%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 3848
      Outlier Samples: 232
      Outlier Percentage: 6.03%
    Class 1:
      Total Samples: 285
      Outlier Samples: 0
      Outlier Percentage: 0.00%
    Class 2:
      Total Samples: 774
      Outlier Samples: 0
      Outlier Percentage: 0.00%
    Class 3:
      Total Samples: 970
      Outlier Samples: 0
      Outlier Percentage: 0.00%

Feature_7:
  Total Outliers: 295 (5.02%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 3848
      Outlier Samples: 259
      Outlier Percentage: 6.73%
    Class 1:
      Total Samples: 285
      Outlier Samples: 26
      Outlier Percentage: 9.12%
    Class 2:
      Total Samples: 774
      Outlier Samples: 53
      Outlier Percentage: 6.85%
    Class 3:
      Total Samples: 970
      Outlier Samples: 2
      Outlier Percentage: 0.21%

Feature_8:
  Total Outliers: 870 (14.80%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 3848
      Outlier Samples: 834
      Outlier Percentage: 21.67%
    Class 1:
      Total Samples: 285
      Outlier Samples: 39
      Outlier Percentage: 13.68%
    Class 2:
      Total Samples: 774
      Outlier Samples: 0
      Outlier Percentage: 0.00%
    Class 3:
      Total Samples: 970
      Outlier Samples: 7
      Outlier Percentage: 0.72%

Feature_9:
  Total Outliers: 1028 (17.49%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 3848
      Outlier Samples: 475
      Outlier Percentage: 12.34%
    Class 1:
      Total Samples: 285
      Outlier Samples: 26
      Outlier Percentage: 9.12%
    Class 2:
      Total Samples: 774
      Outlier Samples: 189
      Outlier Percentage: 24.42%
    Class 3:
      Total Samples: 970
      Outlier Samples: 7
      Outlier Percentage: 0.72%

Feature_10:
  Total Outliers: 406 (6.91%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 3848
      Outlier Samples: 898
      Outlier Percentage: 23.34%
    Class 1:
      Total Samples: 285
      Outlier Samples: 26
      Outlier Percentage: 9.12%
    Class 2:
      Total Samples: 774
      Outlier Samples: 125
      Outlier Percentage: 16.15%
    Class 3:
      Total Samples: 970
      Outlier Samples: 4
      Outlier Percentage: 0.41%

Feature_11:
  Total Outliers: 461 (7.84%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 3848
      Outlier Samples: 369
      Outlier Percentage: 9.59%
    Class 1:
      Total Samples: 285
      Outlier Samples: 29
      Outlier Percentage: 10.18%
    Class 2:
      Total Samples: 774
      Outlier Samples: 193
      Outlier Percentage: 24.94%
    Class 3:
      Total Samples: 970
      Outlier Samples: 143
      Outlier Percentage: 14.74%

Feature_12:
  Total Outliers: 460 (7.83%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 3848
      Outlier Samples: 634
      Outlier Percentage: 16.48%
    Class 1:
      Total Samples: 285
      Outlier Samples: 38
      Outlier Percentage: 13.33%
    Class 2:
      Total Samples: 774
      Outlier Samples: 0
      Outlier Percentage: 0.00%
    Class 3:
      Total Samples: 970
      Outlier Samples: 11
      Outlier Percentage: 1.13%

Feature_13:
  Total Outliers: 1265 (21.52%)
  Class-wise Outlier Breakdown:
    Class 0:
      Total Samples: 3848
      Outlier Samples: 795
      Outlier Percentage: 20.66%
    Class 1:
      Total Samples: 285
      Outlier Samples: 0
      Outlier Percentage: 0.00%
    Class 2:
      Total Samples: 774
      Outlier Samples: 0
      Outlier Percentage: 0.00%
    Class 3:
      Total Samples: 970
      Outlier Samples: 122
      Outlier Percentage: 12.58%

import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
from scipy import stats
import scipy.stats as stats_scipy

def analyze_class_distribution(y_train, y_test, y_val):
    """
    Analyze class distribution across train, test, and validation sets.

    Args:
        y_train (np.ndarray): Training labels.
        y_test (np.ndarray): Test labels.
        y_val (np.ndarray): Validation labels.
    """
    def get_class_distribution(labels):
        unique, counts = np.unique(labels, return_counts=True)
        total = len(labels)
        percentages = (counts / total) * 100
        return dict(zip(unique, percentages))

    print("Train Set Class Distribution:")
    train_dist = get_class_distribution(y_train)
    for cls, percentage in train_dist.items():
        print(f"Class {cls}: {percentage:.2f}%")

    print("\nTest Set Class Distribution:")
    test_dist = get_class_distribution(y_test)
    for cls, percentage in test_dist.items():
        print(f"Class {cls}: {percentage:.2f}%")

    print("\nValidation Set Class Distribution:")
    val_dist = get_class_distribution(y_val)
    for cls, percentage in val_dist.items():
        print(f"Class {cls}: {percentage:.2f}%")

    return train_dist, test_dist, val_dist

distributions = analyze_class_distribution(y_train, y_test, y_val)
print(distributions)

def compare_distributions(train_data, test_data, val_data, feature_name):
    """
    Compare distributions of a specific feature across train, test, and validation sets.

    Parameters:
        train_data: Training dataset feature column.
        test_data: Test dataset feature column.
        val_data: Validation dataset feature column.
        feature_name: Name of the feature being analyzed.

    Returns:
        Dictionary of statistical test results and visualization.
    """
    results = {}

    # 1. Statistical Tests (Kolmogorov-Smirnov Test)
    def ks_test_between_sets(set1, set2, set_names):
        statistic, p_value = stats.ks_2samp(set1, set2)
        return {
            'statistic': statistic,
            'p_value': p_value,
            'sets_compared': set_names,
            'significant_difference': p_value < 0.05
        }

    results['ks_tests'] = {
        'train_vs_test': ks_test_between_sets(train_data, test_data, 'Train vs Test'),
        'train_vs_val': ks_test_between_sets(train_data, val_data, 'Train vs Validation'),
        'test_vs_val': ks_test_between_sets(test_data, val_data, 'Test vs Validation')
    }

    # 2. Distribution Visualization
    plt.figure(figsize=(15, 5))
    plt.subplot(131)
    plt.hist(train_data, bins=30, alpha=0.5, label='Train')
    plt.hist(test_data, bins=30, alpha=0.5, label='Test')
    plt.hist(val_data, bins=30, alpha=0.5, label='Validation')
    plt.title(f'{feature_name} Distribution')
    plt.xlabel(feature_name)
    plt.ylabel('Frequency')
    plt.legend()

    plt.subplot(132)
    box_data = [train_data, test_data, val_data]
    plt.boxplot(box_data, labels=['Train', 'Test', 'Validation'])
    plt.title(f'{feature_name} Box Plot')
    plt.ylabel(feature_name)

    plt.subplot(133)
    sns.kdeplot(train_data, label='Train')
    sns.kdeplot(test_data, label='Test')
    sns.kdeplot(val_data, label='Validation')
    plt.title(f'{feature_name} Density Plot')
    plt.xlabel(feature_name)
    plt.ylabel('Density')
    plt.legend()
    plt.tight_layout()
    plt.show()

    # 3. Distribution Metrics (Jensen-Shannon Divergence)
    def jensen_shannon_divergence(p, q):
        p = np.asarray(p)
        q = np.asarray(q)
        m = 0.5 * (p + q)
        return 0.5 * (stats_scipy.entropy(p, m) + stats_scipy.entropy(q, m))

    def get_distribution_histogram(data, bins=30):
        hist, _ = np.histogram(data, bins=bins, density=True)
        return hist

    results['distribution_metrics'] = {
        'train_vs_test_js_div': jensen_shannon_divergence(
            get_distribution_histogram(train_data),
            get_distribution_histogram(test_data)
        ),
        'train_vs_val_js_div': jensen_shannon_divergence(
            get_distribution_histogram(train_data),
            get_distribution_histogram(val_data)
        ),
        'test_vs_val_js_div': jensen_shannon_divergence(
            get_distribution_histogram(test_data),
            get_distribution_histogram(val_data)
        )
    }
    return results

def analyze_feature_distributions(X_train, X_test, X_val, feature_names=None):
    """
    Analyze distributions for multiple features.

    Parameters:
        X_train: Training dataset (2D NumPy array).
        X_test: Test dataset (2D NumPy array).
        X_val: Validation dataset (2D NumPy array).
        feature_names: Optional list of feature names (default is column indices).
    """
    if feature_names is None:
        feature_names = [f'Feature_{i}' for i in range(X_train.shape[1])]

    assert X_train.shape[1] == len(feature_names), "Number of feature names must match number of columns"
    comprehensive_results = {}
    for idx, feature in enumerate(feature_names):
        print(f"\n--- Analysis for Feature: {feature} ---")
        train_feature = X_train[:, idx]
        test_feature = X_test[:, idx]
        val_feature = X_val[:, idx]
        results = compare_distributions(train_feature, test_feature, val_feature, feature)
        comprehensive_results[feature] = results
        print("Kolmogorov-Smirnov Test Results:")
        for key, test_result in results['ks_tests'].items():
            print(f"{test_result['sets_compared']}:")
            print(f"  Statistic: {test_result['statistic']:.4f}")
            print(f"  P-value: {test_result['p_value']:.4f}")
            print(f"  Significant Difference: {test_result['significant_difference']}")
        print("\nJensen-Shannon Divergence:")
        for key, div_value in results['distribution_metrics'].items():
            print(f"{key}: {div_value:.4f}")
    return comprehensive_results


features = X_train_scaled.columns.tolist()
analyze_feature_distributions(X_train_scaled.to_numpy(), X_test_scaled.to_numpy(), X_val_scaled.to_numpy(), features)

Train Set Class Distribution:
Class 0: 65.48%
Class 1: 4.86%
Class 2: 13.17%
Class 3: 16.50%

Test Set Class Distribution:
Class 0: 65.48%
Class 1: 4.87%
Class 2: 13.15%
Class 3: 16.51%

Validation Set Class Distribution:
Class 0: 65.48%
Class 1: 4.85%
Class 2: 13.17%
Class 3: 16.51%
({np.int64(0): np.float64(65.47929665343165), np.int64(1): np.float64(4.855360181508791), np.int64(2): np.float64(13.16505955757232), np.int64(3): np.float64(16.500283607487237)}, {np.int64(0): np.float64(65.4755827803301), np.int64(1): np.float64(4.8664284498893995), np.int64(2): np.float64(13.152969201973796), np.int64(3): np.float64(16.505019567806702)}, {np.int64(0): np.float64(65.4755827803301), np.int64(1): np.float64(4.849412965798877), np.int64(2): np.float64(13.169984686064318), np.int64(3): np.float64(16.505019567806702)})

--- Analysis for Feature: Flow Duration ---

/tmp/ipython-input-368-813627218.py:86: MatplotlibDeprecationWarning: The 'labels' parameter of boxplot() has been renamed 'tick_labels' since Matplotlib 3.9; support for the old name will be dropped in 3.11.
  plt.boxplot(box_data, labels=['Train', 'Test', 'Validation'])

Kolmogorov-Smirnov Test Results:
Train vs Test:
  Statistic: 0.0075
  P-value: 0.9632
  Significant Difference: False
Train vs Validation:
  Statistic: 0.0136
  P-value: 0.3831
  Significant Difference: False
Test vs Validation:
  Statistic: 0.0133
  P-value: 0.6787
  Significant Difference: False

Jensen-Shannon Divergence:
train_vs_test_js_div: 0.0010
train_vs_val_js_div: 0.0008
test_vs_val_js_div: 0.0008

--- Analysis for Feature: Flow IAT Mean ---

/tmp/ipython-input-368-813627218.py:86: MatplotlibDeprecationWarning: The 'labels' parameter of boxplot() has been renamed 'tick_labels' since Matplotlib 3.9; support for the old name will be dropped in 3.11.
  plt.boxplot(box_data, labels=['Train', 'Test', 'Validation'])

Kolmogorov-Smirnov Test Results:
Train vs Test:
  Statistic: 0.0108
  P-value: 0.6738
  Significant Difference: False
Train vs Validation:
  Statistic: 0.0121
  P-value: 0.5356
  Significant Difference: False
Test vs Validation:
  Statistic: 0.0131
  P-value: 0.6942
  Significant Difference: False

Jensen-Shannon Divergence:
train_vs_test_js_div: 0.0370
train_vs_val_js_div: 0.0391
test_vs_val_js_div: 0.0008

--- Analysis for Feature: Bwd Packet Length Mean ---

/tmp/ipython-input-368-813627218.py:86: MatplotlibDeprecationWarning: The 'labels' parameter of boxplot() has been renamed 'tick_labels' since Matplotlib 3.9; support for the old name will be dropped in 3.11.
  plt.boxplot(box_data, labels=['Train', 'Test', 'Validation'])

Kolmogorov-Smirnov Test Results:
Train vs Test:
  Statistic: 0.0085
  P-value: 0.9050
  Significant Difference: False
Train vs Validation:
  Statistic: 0.0142
  P-value: 0.3313
  Significant Difference: False
Test vs Validation:
  Statistic: 0.0143
  P-value: 0.5857
  Significant Difference: False

Jensen-Shannon Divergence:
train_vs_test_js_div: 0.0713
train_vs_val_js_div: 0.0704
test_vs_val_js_div: 0.0013

--- Analysis for Feature: Bwd Packet Length Max ---

/tmp/ipython-input-368-813627218.py:86: MatplotlibDeprecationWarning: The 'labels' parameter of boxplot() has been renamed 'tick_labels' since Matplotlib 3.9; support for the old name will be dropped in 3.11.
  plt.boxplot(box_data, labels=['Train', 'Test', 'Validation'])

Kolmogorov-Smirnov Test Results:
Train vs Test:
  Statistic: 0.0090
  P-value: 0.8634
  Significant Difference: False
Train vs Validation:
  Statistic: 0.0149
  P-value: 0.2757
  Significant Difference: False
Test vs Validation:
  Statistic: 0.0160
  P-value: 0.4399
  Significant Difference: False

Jensen-Shannon Divergence:
train_vs_test_js_div: 0.0005
train_vs_val_js_div: 0.0013
test_vs_val_js_div: 0.0015

--- Analysis for Feature: Flow Bytes/s ---

/tmp/ipython-input-368-813627218.py:86: MatplotlibDeprecationWarning: The 'labels' parameter of boxplot() has been renamed 'tick_labels' since Matplotlib 3.9; support for the old name will be dropped in 3.11.
  plt.boxplot(box_data, labels=['Train', 'Test', 'Validation'])

import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
from sklearn.decomposition import PCA
from sklearn.manifold import TSNE
from sklearn.preprocessing import StandardScaler


features = df.drop(columns=['Label', 'Destination Port'], errors='ignore')
labels = df['Label']


scaler = StandardScaler()
features_scaled = scaler.fit_transform(features)

#  PCA visualization (2 components)
pca = PCA(n_components=2)
pca_result = pca.fit_transform(features_scaled)
pca_df = pd.DataFrame(data=pca_result, columns=['PC1', 'PC2'])
pca_df['Label'] = labels.values

plt.figure(figsize=(8, 6))
sns.scatterplot(x='PC1', y='PC2', hue='Label', data=pca_df, palette='tab10', alpha=0.7)
plt.title("PCA - 2 Components")
plt.show()

#  t-SNE (2 components)
tsne = TSNE(n_components=2, random_state=42, perplexity=30, max_iter=1000)
tsne_result = tsne.fit_transform(features_scaled)
tsne_df = pd.DataFrame(data=tsne_result, columns=['TSNE1', 'TSNE2'])
tsne_df['Label'] = labels.values

plt.figure(figsize=(8, 6))
sns.scatterplot(x='TSNE1', y='TSNE2', hue='Label', data=tsne_df, palette='tab10', alpha=0.7)
plt.title("t-SNE Visualization")
plt.show()

X_train_tensor = torch.tensor(X_train_scaled.values, dtype=torch.float32)
y_train_tensor = torch.tensor(y_train, dtype=torch.long)
X_val_tensor = torch.tensor(X_val_scaled.values, dtype=torch.float32)
y_val_tensor = torch.tensor(y_val.values, dtype=torch.long)
X_test_tensor = torch.tensor(X_test_scaled.values, dtype=torch.float32)
y_test_tensor = torch.tensor(y_test.values, dtype=torch.long)

train_dataset = TensorDataset(X_train_tensor, y_train_tensor)
val_dataset = TensorDataset(X_val_tensor, y_val_tensor)
test_dataset = TensorDataset(X_test_tensor, y_test_tensor)

train_loader = DataLoader(train_dataset, batch_size=64, shuffle=True)
val_loader = DataLoader(val_dataset, batch_size=64, shuffle=False)
test_loader = DataLoader(test_dataset, batch_size=64, shuffle=False)

#X_train_tensor_SS = torch.tensor(X_train_SS, dtype=torch.float32)
#y_train_tensor_SS = torch.tensor(y_train, dtype=torch.long)
#X_val_tensor_SS = torch.tensor(X_val_SS, dtype=torch.float32)
#y_val_tensor_SS = torch.tensor(y_val.values, dtype=torch.long)
#X_test_tensor_SS = torch.tensor(X_test_SS, dtype=torch.float32)
#y_test_tensor_SS = torch.tensor(y_test.values, dtype=torch.long)

#train_dataset_SS = TensorDataset(X_train_tensor_SS, y_train_tensor_SS)
#val_dataset_SS = TensorDataset(X_val_tensor_SS, y_val_tensor_SS)
#test_dataset_SS = TensorDataset(X_test_tensor_SS, y_test_tensor_SS)

#train_loader_SS = DataLoader(train_dataset_SS, batch_size=64, shuffle=True)
#val_loader_SS = DataLoader(val_dataset_SS, batch_size=64, shuffle=False)
#test_loader_SS = DataLoader(test_dataset_SS, batch_size=64, shuffle=False)

# Shallow NN

class SimpleNN(nn.Module):
    def __init__(self, input_size=14, hidden_size=32, output_size=4):
        super(SimpleNN, self).__init__()
        self.fc1 = nn.Linear(input_size, hidden_size)  # Hidden layer
        self.fc2 = nn.Linear(hidden_size, output_size)  # Output layer

    def forward(self, x):
        x = self.fc1(x)  # Linear activation
        x = self.fc2(x)  # Linear output
        return x

# setting the device to cuda for the GPU
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
#device = "cpu"
print(f"The device is set to: {device}")

The device is set to: cuda

def training_loop(model, train_loader, val_loader, train_dataset, val_dataset, device, optimizer, criterion):
    """
    Train the model for a specified number of epochs and track the training and validation losses.

    Args:
        model: The neural network model to be trained.
        train_loader: DataLoader for the training dataset.
        val_loader: DataLoader for the validation dataset.
        train_dataset: The entire training dataset.
        val_dataset: The entire validation dataset.
        device: Device to run the model on (CPU or GPU).
        optimizer: The optimization algorithm to use for updating model parameters.
        criterion: The loss function to use for computing the loss.

    Returns:
        None: The function plots the training and validation loss over the epochs.
    """
    # Move the model to the correct device before training
    model.to(device)  # Add this line to move the model to the device

    # Record the start time
    start_time = time.time()

    # Define training parameters
    # we set patience to 30
    num_epochs = 100
    patience = 30
    patience_counter = 0
    best_loss = -1

    train_losses = []
    val_losses = []

    # Training loop
    for epoch in range(num_epochs):  # Train for epochs
        train_loss = 0
        val_loss = 0
        model.train()  # Set model to training mode
        for batch_X, batch_y in train_loader:
            batch_X, batch_y = batch_X.to(device), batch_y.to(device)  # Move batch to GPU
            optimizer.zero_grad()  # Clear previous gradients
            outputs = model(batch_X)  # Forward pass
            loss = criterion(outputs, batch_y)  # Compute loss
            loss.backward()  # Backpropagation
            optimizer.step()  # Update model parameters
            train_loss += loss.item() * batch_X.size(0)
        train_loss /= len(train_dataset)
        train_losses.append(train_loss)  # Store training loss

        # Compute Validation loss during training
        model.eval()  # Set model to evaluation mode
        with torch.no_grad():  # Disable gradient computation for validation
            for batch_X, batch_y in val_loader:
                batch_X, batch_y = batch_X.to(device), batch_y.to(device)  # Move batch to GPU
                val_outputs = model(batch_X)  # Forward pass
                loss = criterion(val_outputs, batch_y)  # Compute validation loss
                val_loss += loss.item() * batch_X.size(0)
            val_loss /= len(val_dataset)
            val_losses.append(val_loss)  # Store validation loss
            # Early stopping condition
        if val_loss <= best_loss:
            best_loss = val_loss
            patience_counter = 0
        else:
            patience_counter += 1
        if patience_counter >= patience:
            print(f"Early stopping triggered at epoch number {epoch} ")
            break

        if (epoch + 1) % 20 == 0:
            print(f'Epoch {epoch+1}/{num_epochs}, Train Loss: {train_losses[-1]:.4f}, Val Loss: {val_losses[-1]:.4f}')


    # Record the end time
    end_time = time.time()
    # Calculate the elapsed time
    elapsed_time = end_time - start_time
    print(f'The function took {elapsed_time:.4f} seconds to execute.')

    # Plot the training and validation loss
    plt.figure(figsize=(10, 5))
    plt.plot(train_losses, label='Train Loss')
    plt.plot(val_losses, label='Validation Loss')
    plt.xlabel('Epoch')
    plt.ylabel('Loss')
    plt.title('Training and Validation Loss')
    plt.legend()
    plt.show()

def plot_confusion_matrix(y_true, y_pred, class_names):
    """
    Creates and displays a confusion matrix to evaluate classification performance.

    Args:
        y_true: Ground truth labels
        y_pred: Predicted labels
        class_names: List of class names for axis labels
    """
    cm = confusion_matrix(y_true, y_pred)  # Compute confusion matrix

    plt.figure(figsize=(8, 6))
    sns.heatmap(cm, annot=True, fmt="d", cmap="Blues", xticklabels=class_names, yticklabels=class_names)

    plt.xlabel("Predicted Label")
    plt.ylabel("True Label")
    plt.title("Confusion Matrix")
    plt.show()
import torch
import time
import numpy as np
from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score, classification_report

def testing_model(model, dataloader, device):
    """
    Evaluate the model on a given dataloader and compute accuracy, precision, recall, and F1-score.

    Args:
        model: The trained model.
        dataloader: DataLoader for the dataset.
        device: Device to run the model on (CPU or GPU).

    Returns:
        dict: Dictionary containing accuracy, precision, recall, and F1-score.
    """
    # Record the start time
    start_time = time.time()

    model.eval()  # Set the model to evaluation mode
    all_labels = []
    all_predictions = []

    with torch.no_grad():  # Disable gradient computation for efficiency
        for inputs, labels in dataloader:
            inputs, labels = inputs.to(device), labels.to(device)  # Move batch to GPU
            outputs = model(inputs)
            _, predicted = torch.max(outputs, 1)  # Get the class with highest score
            all_labels.extend(labels.cpu().numpy())
            all_predictions.extend(predicted.cpu().numpy())

    # Compute metrics
    accuracy = accuracy_score(all_labels, all_predictions) * 100
    precision = precision_score(all_labels, all_predictions, average=None, zero_division=0)
    recall = recall_score(all_labels, all_predictions, average=None, zero_division=0)
    f1 = f1_score(all_labels, all_predictions, average=None, zero_division=0)

    # Generate classification report
    class_names = [f"Class {cls}" for cls in np.unique(all_labels)]
    print("\nClassification Report:")
    print(classification_report(all_labels, all_predictions, target_names=class_names, zero_division=0))

    # Generation of confusion matrix
    plot_confusion_matrix(all_labels, all_predictions, class_names)

    # Record the end time
    end_time = time.time()
    elapsed_time = end_time - start_time
    print(f'The function took {elapsed_time:.4f} seconds to execute.')

    # Return results as a dictionary
    return {
        "accuracy": accuracy,
        "precision": precision.tolist(),
        "recall": recall.tolist(),
        "f1_score": f1.tolist()
    }

# model with a hidden size of 32 nodes
n_input_nodes = 14
criterion = nn.CrossEntropyLoss()
model32 = SimpleNN(n_input_nodes,32)
optimizer = optim.AdamW(model32.parameters(), lr=0.0005)
print(model32)
print(criterion)
print(optimizer)

SimpleNN(
  (fc1): Linear(in_features=14, out_features=32, bias=True)
  (fc2): Linear(in_features=32, out_features=4, bias=True)
)
CrossEntropyLoss()
AdamW (
Parameter Group 0
    amsgrad: False
    betas: (0.9, 0.999)
    capturable: False
    differentiable: False
    eps: 1e-08
    foreach: None
    fused: None
    lr: 0.0005
    maximize: False
    weight_decay: 0.01
)

training_loop(model32, train_loader, val_loader, train_dataset, val_dataset, device, optimizer, criterion)

Epoch 20/100, Train Loss: 0.3317, Val Loss: 0.3372
Early stopping triggered at epoch number 29 
The function took 17.7546 seconds to execute.

train_accuracy = testing_model(model32,train_loader,device)
val_accuracy = testing_model(model32,val_loader,device)
test_accuracy = testing_model(model32,test_loader,device)

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.93      0.94      0.94     11544
     Class 1       0.83      0.65      0.73       856
     Class 2       0.98      0.86      0.91      2321
     Class 3       0.81      0.91      0.86      2909

    accuracy                           0.91     17630
   macro avg       0.89      0.84      0.86     17630
weighted avg       0.91      0.91      0.91     17630

The function took 0.7791 seconds to execute.

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.93      0.94      0.93      3848
     Class 1       0.77      0.65      0.70       285
     Class 2       0.98      0.83      0.90       774
     Class 3       0.81      0.91      0.86       970

    accuracy                           0.91      5877
   macro avg       0.87      0.83      0.85      5877
weighted avg       0.91      0.91      0.91      5877

The function took 0.2987 seconds to execute.

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.93      0.94      0.94      3848
     Class 1       0.82      0.65      0.72       286
     Class 2       0.97      0.87      0.92       773
     Class 3       0.81      0.91      0.86       970

    accuracy                           0.91      5877
   macro avg       0.88      0.84      0.86      5877
weighted avg       0.91      0.91      0.91      5877

The function took 0.3050 seconds to execute.

n_input_nodes = 14
criterion = nn.CrossEntropyLoss()
model64 = SimpleNN(input_size=n_input_nodes,hidden_size=64)
optimizer = optim.AdamW(model64.parameters(), lr=0.0005)
print(model64)
print(criterion)
print(optimizer)

training_loop(model64, train_loader, val_loader, train_dataset, val_dataset, device, optimizer, criterion)

SimpleNN(
  (fc1): Linear(in_features=14, out_features=64, bias=True)
  (fc2): Linear(in_features=64, out_features=4, bias=True)
)
CrossEntropyLoss()
AdamW (
Parameter Group 0
    amsgrad: False
    betas: (0.9, 0.999)
    capturable: False
    differentiable: False
    eps: 1e-08
    foreach: None
    fused: None
    lr: 0.0005
    maximize: False
    weight_decay: 0.01
)
Epoch 20/100, Train Loss: 0.3234, Val Loss: 0.3253
Early stopping triggered at epoch number 29 
The function took 17.2425 seconds to execute.

train_accuracy = testing_model(model64,train_loader,device)
val_accuracy = testing_model(model64,val_loader,device)
test_accuracy = testing_model(model64,test_loader,device)

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.89      0.96      0.92     11544
     Class 1       0.00      0.00      0.00       856
     Class 2       0.98      0.86      0.92      2321
     Class 3       0.83      0.89      0.86      2909

    accuracy                           0.89     17630
   macro avg       0.67      0.68      0.67     17630
weighted avg       0.85      0.89      0.86     17630

The function took 0.5171 seconds to execute.

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.88      0.96      0.92      3848
     Class 1       0.00      0.00      0.00       285
     Class 2       0.99      0.84      0.91       774
     Class 3       0.83      0.90      0.86       970

    accuracy                           0.88      5877
   macro avg       0.68      0.67      0.67      5877
weighted avg       0.85      0.88      0.86      5877

The function took 0.2998 seconds to execute.

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.89      0.95      0.92      3848
     Class 1       0.00      0.00      0.00       286
     Class 2       0.98      0.87      0.92       773
     Class 3       0.82      0.89      0.85       970

    accuracy                           0.89      5877
   macro avg       0.67      0.68      0.67      5877
weighted avg       0.85      0.89      0.86      5877

The function took 0.6550 seconds to execute.

n_input_nodes = 14
criterion = nn.CrossEntropyLoss()
model_128 = SimpleNN(input_size=n_input_nodes,hidden_size=128)
optimizer = optim.AdamW(model_128.parameters(), lr=0.0005)
print(model_128)
print(criterion)
print(optimizer)

training_loop(model_128, train_loader, val_loader, train_dataset, val_dataset, device, optimizer, criterion)

SimpleNN(
  (fc1): Linear(in_features=14, out_features=128, bias=True)
  (fc2): Linear(in_features=128, out_features=4, bias=True)
)
CrossEntropyLoss()
AdamW (
Parameter Group 0
    amsgrad: False
    betas: (0.9, 0.999)
    capturable: False
    differentiable: False
    eps: 1e-08
    foreach: None
    fused: None
    lr: 0.0005
    maximize: False
    weight_decay: 0.01
)
Epoch 20/100, Train Loss: 0.3175, Val Loss: 0.3210
Early stopping triggered at epoch number 29 
The function took 17.7467 seconds to execute.

train_accuracy = testing_model(model_128,train_loader,device)
val_accuracy = testing_model(model_128,val_loader,device)
test_accuracy = testing_model(model_128,test_loader,device)

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.89      0.96      0.92     11544
     Class 1       0.00      0.00      0.00       856
     Class 2       0.99      0.86      0.92      2321
     Class 3       0.82      0.89      0.86      2909

    accuracy                           0.89     17630
   macro avg       0.68      0.68      0.67     17630
weighted avg       0.85      0.89      0.87     17630

The function took 0.8446 seconds to execute.

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.89      0.96      0.92      3848
     Class 1       0.00      0.00      0.00       285
     Class 2       0.99      0.84      0.91       774
     Class 3       0.83      0.90      0.86       970

    accuracy                           0.89      5877
   macro avg       0.68      0.67      0.67      5877
weighted avg       0.85      0.89      0.86      5877

The function took 0.4465 seconds to execute.

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.89      0.95      0.92      3848
     Class 1       0.00      0.00      0.00       286
     Class 2       0.98      0.87      0.92       773
     Class 3       0.82      0.89      0.86       970

    accuracy                           0.89      5877
   macro avg       0.67      0.68      0.67      5877
weighted avg       0.85      0.89      0.87      5877

The function took 0.4665 seconds to execute.

class SimpleNN_ReLU(nn.Module):
    def __init__(self, input_size=14, hidden_size=32, output_size=4):
        super(SimpleNN_ReLU, self).__init__()
        self.fc1 = nn.Linear(input_size, hidden_size)  # Hidden layer
        self.fc2 = nn.Linear(hidden_size, output_size)  # Output layer
        self.relu = nn.ReLU()

    def forward(self, x):
        x = self.relu(self.fc1(x))
        x = self.fc2(x)
        return x

relu_nn = SimpleNN_ReLU(input_size=14,hidden_size=64)
optimizer = optim.AdamW(relu_nn.parameters(), lr=0.0005)
criterion = nn.CrossEntropyLoss()
print(relu_nn)
print(criterion)
print(optimizer)

SimpleNN_ReLU(
  (fc1): Linear(in_features=14, out_features=64, bias=True)
  (fc2): Linear(in_features=64, out_features=4, bias=True)
  (relu): ReLU()
)
CrossEntropyLoss()
AdamW (
Parameter Group 0
    amsgrad: False
    betas: (0.9, 0.999)
    capturable: False
    differentiable: False
    eps: 1e-08
    foreach: None
    fused: None
    lr: 0.0005
    maximize: False
    weight_decay: 0.01
)

training_loop(relu_nn, train_loader, val_loader, train_dataset, val_dataset, device, optimizer, criterion)

Epoch 20/100, Train Loss: 0.1629, Val Loss: 0.1834
Early stopping triggered at epoch number 29 
The function took 16.8850 seconds to execute.

train_accuracy = testing_model(relu_nn,train_loader,device)
val_accuracy = testing_model(relu_nn,val_loader,device)
test_accuracy = testing_model(relu_nn,test_loader,device)

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.95      0.97      0.96     11544
     Class 1       0.80      0.89      0.84       856
     Class 2       1.00      0.87      0.93      2321
     Class 3       0.94      0.92      0.93      2909

    accuracy                           0.95     17630
   macro avg       0.92      0.92      0.92     17630
weighted avg       0.95      0.95      0.95     17630

The function took 0.7446 seconds to execute.

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.95      0.97      0.96      3848
     Class 1       0.77      0.91      0.83       285
     Class 2       0.99      0.85      0.92       774
     Class 3       0.93      0.92      0.93       970

    accuracy                           0.94      5877
   macro avg       0.91      0.91      0.91      5877
weighted avg       0.94      0.94      0.94      5877

The function took 0.2843 seconds to execute.

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.95      0.97      0.96      3848
     Class 1       0.78      0.89      0.83       286
     Class 2       1.00      0.88      0.94       773
     Class 3       0.93      0.92      0.93       970

    accuracy                           0.95      5877
   macro avg       0.92      0.92      0.91      5877
weighted avg       0.95      0.95      0.95      5877

The function took 0.2754 seconds to execute.

# here we take the test test, we replace port 80 with port 8080 (so we don't have to encode the port)
# i am reloading the dataset but without retraining the model
X_test_8080 = X_test.copy()
X_test_8080 = X_test_8080.assign(Label=y_test.values)  # Append y_test as a column named 'Label'

# Replaced port 80 with 8080 where Label is 1
X_test_8080.loc[(X_test_8080['Label'] == 1) & (X_test_8080['Destination Port'] == 80), 'Destination Port'] = 8080

X_test_8080 = X_test_8080.drop(columns=['Label'])

# 3. Preprocessing the test set
X_test_8080.drop_duplicates(inplace=True)
# For the test set, we applied the same log transform and the same scaler fitted on the training set




X_test_scaled = scaler.transform(X_test_8080)
X_test_scaled = pd.DataFrame(X_test_scaled, columns=X_test_8080.columns, index=X_test_8080.index)


X_test_tensor = torch.tensor(X_test_scaled.values, dtype=torch.float32)
y_test_tensor = torch.tensor(y_test.values, dtype=torch.long)

test_dataset = TensorDataset(X_test_tensor, y_test_tensor)

test_loader = DataLoader(test_dataset, batch_size=64, shuffle=False)

test_accuracy = testing_model(relu_nn,test_loader,device)

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.90      0.97      0.93      3848
     Class 1       0.20      0.06      0.09       286
     Class 2       1.00      0.88      0.94       773
     Class 3       0.93      0.92      0.93       970

    accuracy                           0.91      5877
   macro avg       0.76      0.71      0.72      5877
weighted avg       0.88      0.91      0.89      5877

The function took 0.2859 seconds to execute.

downloaded.GetContentFile('data_set_lab_1.csv')  # Download to a local file
df = pd.read_csv('data_set_lab_1.csv')


# label encoding
label_encoder = LabelEncoder()
df['Label'] = label_encoder.fit_transform(df['Label'])

label_mapping = dict(zip(label_encoder.classes_, label_encoder.transform(label_encoder.classes_)))


print(label_mapping)


# {'Benign': np.int64(0), 'Brute Force': np.int64(1), 'DoS Hulk': np.int64(2), 'PortScan': np.int64(3)}

# pre processing
# We can decide what to do, since there are not many rows that have inf as value we can simply decide to drop them (even if they are form a class that is not really present in the dataset)
df.replace([np.inf, -np.inf], np.nan, inplace=True)
df.dropna(inplace=True)

# Convert relevant columns to numeric before comparison
for column in df.select_dtypes(include=['object']).columns:
    if column not in ['Label']:
        try:
            df[column] = pd.to_numeric(df[column])
        except ValueError:
            print(f"Could not convert column '{column}' to numeric.")

#remove of negative numbers
df = df[(df >= 0).all(axis=1)]

# Reset the index if needed
df = df.reset_index(drop=True)



# remove the biased feature
df = df.drop(columns=['Destination Port'])



cols_to_drop = [
    'Fwd PSH Flags',
    'Subflow Fwd Packets'

]

df = df.drop(columns=cols_to_drop)

df.drop_duplicates(inplace=True)





import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import FunctionTransformer, RobustScaler
import matplotlib.pyplot as plt
import seaborn as sns


X = df.drop(columns=['Label'])
y = df['Label']

# 1. Split the dataset into training (60%), validation (20%), and test (20%) sets

X_temp, X_test, y_temp, y_test = train_test_split(
    X, y, test_size=0.2, random_state=42, stratify=y
)
X_train, X_val, y_train, y_val = train_test_split(
    X_temp, y_temp, test_size=0.2, random_state=42, stratify=y_temp
)




scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train)
X_val_scaled = scaler.transform(X_val)


X_train_scaled = pd.DataFrame(X_train_scaled, columns=X_train.columns, index=X_train.index)
X_val_scaled = pd.DataFrame(X_val_scaled, columns=X_val.columns, index=X_val.index)


X_test_scaled= scaler.transform(X_test)
X_test_scaled = pd.DataFrame(X_test_scaled, columns=X_test.columns, index=X_test.index)

# This methodology ensures that the FFNN model is trained and validated on pre-processed data in a consistent manner
# and without information leakage from the test or validation sets.

X_train_tensor = torch.tensor(X_train_scaled.values, dtype=torch.float32)
y_train_tensor = torch.tensor(y_train, dtype=torch.long)
X_val_tensor = torch.tensor(X_val_scaled.values, dtype=torch.float32)
y_val_tensor = torch.tensor(y_val.values, dtype=torch.long)
X_test_tensor = torch.tensor(X_test_scaled.values, dtype=torch.float32)
y_test_tensor = torch.tensor(y_test.values, dtype=torch.long)

train_dataset = TensorDataset(X_train_tensor, y_train_tensor)
val_dataset = TensorDataset(X_val_tensor, y_val_tensor)
test_dataset = TensorDataset(X_test_tensor, y_test_tensor)

train_loader = DataLoader(train_dataset, batch_size=64, shuffle=True)
val_loader = DataLoader(val_dataset, batch_size=64, shuffle=False)
test_loader = DataLoader(test_dataset, batch_size=64, shuffle=False)

{'Benign': np.int64(0), 'Brute Force': np.int64(1), 'DoS Hulk': np.int64(2), 'PortScan': np.int64(3)}

relu13f_nn = SimpleNN_ReLU(input_size=13,hidden_size=64)
optimizer = optim.AdamW(relu13f_nn.parameters(), lr=0.0005)
criterion = nn.CrossEntropyLoss()
print(relu13f_nn)
print(criterion)
print(optimizer)

SimpleNN_ReLU(
  (fc1): Linear(in_features=13, out_features=64, bias=True)
  (fc2): Linear(in_features=64, out_features=4, bias=True)
  (relu): ReLU()
)
CrossEntropyLoss()
AdamW (
Parameter Group 0
    amsgrad: False
    betas: (0.9, 0.999)
    capturable: False
    differentiable: False
    eps: 1e-08
    foreach: None
    fused: None
    lr: 0.0005
    maximize: False
    weight_decay: 0.01
)

training_loop(relu13f_nn, train_loader, val_loader, train_dataset, val_dataset, device, optimizer, criterion)

Epoch 20/100, Train Loss: 0.1659, Val Loss: 0.1646
Early stopping triggered at epoch number 29 
The function took 14.5526 seconds to execute.

test_accuracy = testing_model(relu13f_nn,test_loader,device)

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.94      0.98      0.96      3378
     Class 1       0.79      0.89      0.84       285
     Class 2       0.99      0.85      0.91       774
     Class 3       0.57      0.07      0.12        57

    accuracy                           0.94      4494
   macro avg       0.82      0.70      0.71      4494
weighted avg       0.94      0.94      0.93      4494

The function took 0.2766 seconds to execute.

import numpy as np

def analyze_class_distribution(y_train, y_test, y_val):
    """
    Analyze class distribution across train, test, and validation sets

    Args:
        y_train (np.ndarray): Training labels
        y_test (np.ndarray): Test labels
        y_val (np.ndarray): Validation labels
    """
    # Function to calculate class distribution
    def get_class_distribution(labels):
        unique, counts = np.unique(labels, return_counts=True)
        total = len(labels)
        percentages = (counts / total) * 100
        return dict(zip(unique, percentages))

    # Print distributions
    print("Train Set Class Distribution:")
    train_dist = get_class_distribution(y_train)
    for cls, percentage in train_dist.items():
        print(f"Class {cls}: {percentage:.2f}%")

    print("\nTest Set Class Distribution:")
    test_dist = get_class_distribution(y_test)
    for cls, percentage in test_dist.items():
        print(f"Class {cls}: {percentage:.2f}%")

    print("\nValidation Set Class Distribution:")
    val_dist = get_class_distribution(y_val)
    for cls, percentage in val_dist.items():
        print(f"Class {cls}: {percentage:.2f}%")

    return train_dist, test_dist, val_dist





distributions = analyze_class_distribution(y_train, y_test, y_val)
print(distributions)

Train Set Class Distribution:
Class 0: 75.17%
Class 1: 6.36%
Class 2: 17.21%
Class 3: 1.27%

Test Set Class Distribution:
Class 0: 75.17%
Class 1: 6.34%
Class 2: 17.22%
Class 3: 1.27%

Validation Set Class Distribution:
Class 0: 75.16%
Class 1: 6.34%
Class 2: 17.22%
Class 3: 1.28%
({np.int64(0): np.float64(75.16517143055846), np.int64(1): np.float64(6.356492106544266), np.int64(2): np.float64(17.21260171082829), np.int64(3): np.float64(1.2657347520689894)}, {np.int64(0): np.float64(75.16688918558077), np.int64(1): np.float64(6.341789052069426), np.int64(2): np.float64(17.222963951935917), np.int64(3): np.float64(1.268357810413885)}, {np.int64(0): np.float64(75.15994436717664), np.int64(1): np.float64(6.342141863699583), np.int64(2): np.float64(17.21835883171071), np.int64(3): np.float64(1.2795549374130737)})

from sklearn.utils.class_weight import compute_class_weight

# Calculate class weights
class_weights = compute_class_weight(class_weight='balanced', classes=np.unique(y_train), y=y_train)


# Convert to tensor and move to device
class_weights = torch.tensor(class_weights, dtype=torch.float).to(device)



relu13f_nn_cw = SimpleNN_ReLU(input_size=13,hidden_size=64)
optimizer = optim.AdamW(relu13f_nn_cw.parameters(), lr=0.0005)
# Updated criterion with weights
criterion = nn.CrossEntropyLoss(weight = class_weights)
print(relu13f_nn_cw)
print(criterion)
print(optimizer)
print(class_weights)

SimpleNN_ReLU(
  (fc1): Linear(in_features=13, out_features=64, bias=True)
  (fc2): Linear(in_features=64, out_features=4, bias=True)
  (relu): ReLU()
)
CrossEntropyLoss()
AdamW (
Parameter Group 0
    amsgrad: False
    betas: (0.9, 0.999)
    capturable: False
    differentiable: False
    eps: 1e-08
    foreach: None
    fused: None
    lr: 0.0005
    maximize: False
    weight_decay: 0.01
)
tensor([ 0.3326,  3.9330,  1.4524, 19.7514], device='cuda:0')

training_loop(relu13f_nn_cw, train_loader, val_loader, train_dataset, val_dataset, device, optimizer, criterion)

Epoch 20/100, Train Loss: 0.2861, Val Loss: 0.2981
Early stopping triggered at epoch number 29 
The function took 14.3153 seconds to execute.

test =testing_model(relu13f_nn_cw,test_loader,device)

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.97      0.91      0.94      3378
     Class 1       0.72      0.93      0.82       285
     Class 2       0.98      0.88      0.93       774
     Class 3       0.19      0.86      0.31        57

    accuracy                           0.91      4494
   macro avg       0.72      0.90      0.75      4494
weighted avg       0.95      0.91      0.92      4494

The function took 0.2823 seconds to execute.

# ---------------- reproducibility & helpers ----------------
import time, copy, itertools, torch, numpy as np
import torch.nn as nn, torch.optim as optim
from sklearn.metrics import accuracy_score, f1_score

DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
torch.manual_seed(42);  np.random.seed(42)

# ---------------- training_loop + live-plot --------------------------
import time, copy, matplotlib.pyplot as plt

def training_loop(model, train_loader, val_loader,
                  train_ds,   val_ds, device,
                  optimizer, criterion,
                  max_epochs=100, patience=30,
                  verbose_every=20, cfg_name="run"):
    """
      Train a PyTorch model with early-stopping and live loss plotting.

      Args:
          model (torch.nn.Module): Network to train.
          train_loader (DataLoader): Mini-batches for training.
          val_loader (DataLoader): Mini-batches for validation.
          train_ds (Dataset): Full training set (for loss normalisation).
          val_ds (Dataset): Full validation set.
          device (torch.device): CPU or GPU on which to run.
          optimizer (torch.optim.Optimizer): Parameter update rule.
          criterion (Callable): Loss function.
          max_epochs (int, optional):  Default = 100.
          patience (int, optional): Default = 30.
          verbose_every (int, optional): Print progress every N epochs; Default = 20.
          cfg_name (str, optional): Label shown in logs and plot title. Default = "run".

      Returns:
          Tuple[Dict[str, List[float]], torch.nn.Module]:
              • History of training and validation losses.
              • Model restored to the weights that achieved the best validation loss.

    """


    model = model.to(device)
    best_w      = copy.deepcopy(model.state_dict())
    best_loss   = float('inf')
    wait        = 0
    hist_tr, hist_val = [], []
    t0 = time.time()

    for epoch in range(max_epochs):
        # ---------- TRAIN ----------
        model.train()
        running = 0.0
        for xb, yb in train_loader:
            xb, yb = xb.to(device), yb.to(device)
            optimizer.zero_grad()
            loss = criterion(model(xb), yb)
            loss.backward(); optimizer.step()
            running += loss.item() * xb.size(0)
        tr_loss = running / len(train_ds)
        hist_tr.append(tr_loss)

        # ---------- VALID ----------
        model.eval(); running = 0.0
        with torch.no_grad():
            for xb, yb in val_loader:
                xb, yb = xb.to(device), yb.to(device)
                running += criterion(model(xb), yb).item() * xb.size(0)
        val_loss = running / len(val_ds)
        hist_val.append(val_loss)

        # ---------- EARLY-STOP ----------
        if val_loss < best_loss:
            best_loss = val_loss
            best_w    = copy.deepcopy(model.state_dict())
            wait      = 0
        else:
            wait += 1
            if wait >= patience:
                print(f"[{cfg_name}] early-stop @ epoch {epoch:03d}")
                break

        if verbose_every and (epoch+1) % verbose_every == 0:
            print(f"[{cfg_name}] ep {epoch+1:03d}  tr={tr_loss:.4f}  val={val_loss:.4f}")

    # restore best weights to
    model.load_state_dict(best_w)
    print(f"[{cfg_name}] done in {time.time()-t0:.1f}s  best-val={best_loss:.4f}")

    # ---------- PLOT ----------
    plt.figure(figsize=(6,3))
    plt.plot(hist_tr, label='train')
    plt.plot(hist_val,label='val')
    plt.xlabel('epoch'); plt.ylabel('loss'); plt.title(cfg_name); plt.legend()
    plt.tight_layout(); plt.show()

    return {'train': hist_tr, 'val': hist_val}, model

# ---------------- PlainFFNN builder ------------------------
class PlainFFNN(nn.Module):

    """
    Simple feed-forward neural network.

    Args:
        layer_sizes (List[int]): Number of units in each layer, e.g. [13, 32, 8, 4].

    Architecture:
        • Fully connected layers defined by `layer_sizes`
        • ReLU applied to every hidden layer
        • Last layer left linear, returning raw logits

    """
    def __init__(self, layer_sizes):
        super().__init__()
        self.fcs = nn.ModuleList(
            [nn.Linear(i, o) for i, o in zip(layer_sizes[:-1], layer_sizes[1:])]
        )
        self.act = nn.ReLU()

    def forward(self, x):
        for lyr in self.fcs[:-1]:
            x = self.act(lyr(x))
        return self.fcs[-1](x)

# ---------------- architecture sweep -------------------------------
# Run a small hyper-parameter sweep over fully connected architectures:
#   1. Build a candidate grid of 2-5 hidden layers whose widths monotonically
#      decrease (e.g. 32 -> 16 -> 8).  Each architecture is encoded as a list like
#      [input_dim, 32, 16, 8, n_classes].
#   2. For every architecture:
#        • Instantiate a PlainFFNN with ReLU activations
#        • Train it with AdamW + Cross-Entropy (class-weighted), early-stopping
#          max 50 epochs / 15 patience via `training_loop`
#        • Record accuracy on train, validation and test sets using
#          `testing_model`
#   3. Collect all results in a dict -> transpose to a DataFrame for comparison.

import itertools, torch.nn as nn, torch.optim as optim, numpy as np

input_dim  = X_train_tensor.shape[1]
n_classes  = len(torch.unique(y_train_tensor))

def PlainFFNN(sizes):
    layers=[]
    for a,b in zip(sizes[:-1], sizes[1:]):
        layers += [nn.Linear(a,b), nn.ReLU()]
    layers.pop()                        # keep last layer linear (logits)
    return nn.Sequential(*layers)

arch_grid = []
for d in range(2,6):                                      # 2 to 5 hidden layer
    for widths in itertools.product([32,16,8,4,2], repeat=d):
        if list(widths)==sorted(widths, reverse=True):  # enforce descending
            arch_grid.append([input_dim,*widths,n_classes])

results_arch = {}   # {cfg_name : {'train':…, 'val':…, 'test':…}}

for idx, arch in enumerate(arch_grid,1):
    cfg = f"ARCH#{idx:03d}-{arch[1:-1]}"
    net = PlainFFNN(arch)
    opt = optim.AdamW(net.parameters(), lr=5e-4)
    hist, net = training_loop(net, train_loader, val_loader,
                              train_dataset, val_dataset, DEVICE,
                              opt, nn.CrossEntropyLoss(weight=class_weights),
                              max_epochs=50, patience=15,
                              verbose_every=0, cfg_name=cfg)

    tr  = testing_model(net, train_loader, DEVICE)
    vl  = testing_model(net,  val_loader, DEVICE)
    tst = testing_model(net,  test_loader, DEVICE)
    results_arch[cfg] = {'train':tr, 'val':vl, 'test':tst}


pd.DataFrame(results_arch).T

[ARCH#001-[32, 32]] done in 26.6s  best-val=0.2063

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.99      0.92      0.96     10808
     Class 1       0.76      0.96      0.85       914
     Class 2       0.88      0.96      0.92      2475
     Class 3       0.37      0.88      0.53       182

    accuracy                           0.93     14379
   macro avg       0.75      0.93      0.81     14379
weighted avg       0.95      0.93      0.94     14379

The function took 0.4685 seconds to execute.

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.99      0.92      0.95      2702
     Class 1       0.76      0.93      0.84       228
     Class 2       0.87      0.97      0.91       619
     Class 3       0.35      0.87      0.50        46

    accuracy                           0.93      3595
   macro avg       0.74      0.92      0.80      3595
weighted avg       0.95      0.93      0.93      3595

The function took 0.2529 seconds to execute.

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.98      0.93      0.96      3378
     Class 1       0.76      0.94      0.84       285
     Class 2       0.89      0.94      0.91       774
     Class 3       0.37      0.88      0.52        57

    accuracy                           0.93      4494
   macro avg       0.75      0.92      0.81      4494
weighted avg       0.95      0.93      0.94      4494

The function took 0.2838 seconds to execute.
[ARCH#002-[32, 16]] done in 26.7s  best-val=0.2292

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.99      0.89      0.94     10808
     Class 1       0.75      0.96      0.84       914
     Class 2       0.81      0.95      0.87      2475
     Class 3       0.32      0.88      0.46       182

    accuracy                           0.91     14379
   macro avg       0.72      0.92      0.78     14379
weighted avg       0.93      0.91      0.92     14379

The function took 0.4762 seconds to execute.

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.99      0.90      0.94      2702
     Class 1       0.77      0.93      0.84       228
     Class 2       0.81      0.97      0.88       619
     Class 3       0.31      0.89      0.46        46

    accuracy                           0.91      3595
   macro avg       0.72      0.92      0.78      3595
weighted avg       0.94      0.91      0.92      3595

The function took 0.2711 seconds to execute.

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.98      0.90      0.94      3378
     Class 1       0.75      0.93      0.83       285
     Class 2       0.82      0.94      0.88       774
     Class 3       0.32      0.89      0.47        57

    accuracy                           0.91      4494
   macro avg       0.72      0.92      0.78      4494
weighted avg       0.93      0.91      0.92      4494

The function took 0.2763 seconds to execute.
[ARCH#003-[32, 8]] done in 26.4s  best-val=0.2200

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.99      0.90      0.94     10808
     Class 1       0.75      0.96      0.84       914
     Class 2       0.84      0.95      0.89      2475
     Class 3       0.28      0.88      0.42       182

    accuracy                           0.91     14379
   macro avg       0.71      0.92      0.77     14379
weighted avg       0.94      0.91      0.92     14379

class_weights = class_weights.to(DEVICE)
input_dim  = X_train_tensor.shape[1]
n_classes  = len(torch.unique(y_train_tensor))
ref_arch = [input_dim, 32, 32, 32, n_classes]


# ---------------- batch-size sweep ---------------------------------
# Purpose:
#   Evaluate how different mini-batch sizes affect model performance.

batch_grid = [1,32,64,128,512]
bs_scores  = {}

for bs in batch_grid:
    tr_loader = DataLoader(train_dataset,batch_size=bs,shuffle=True)
    vl_loader = DataLoader(val_dataset,  batch_size=bs,shuffle=False)

    net = PlainFFNN(ref_arch)
    opt = optim.AdamW(net.parameters(), lr=5e-4)
    training_loop(net, tr_loader, vl_loader,
                  train_dataset, val_dataset, DEVICE,
                  opt, nn.CrossEntropyLoss(weight=class_weights),
                  max_epochs=50, patience=15,
                  verbose_every=0, cfg_name=f"BS={bs}")

    bs_scores[bs] = {
        'train_acc': testing_model(net, tr_loader, DEVICE),
        'val_acc'  : testing_model(net, vl_loader, DEVICE),
        'test_acc' : testing_model(net, test_loader,DEVICE)
    }

pd.DataFrame(bs_scores).T

[BS=1] done in 1376.0s  best-val=0.1102

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.96      0.98      0.97     10808
     Class 1       0.81      0.95      0.87       914
     Class 2       1.00      0.89      0.94      2475
     Class 3       0.38      0.14      0.21       182

    accuracy                           0.95     14379
   macro avg       0.79      0.74      0.75     14379
weighted avg       0.95      0.95      0.95     14379

The function took 6.9989 seconds to execute.

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.96      0.98      0.97      2702
     Class 1       0.83      0.93      0.88       228
     Class 2       1.00      0.91      0.95       619
     Class 3       0.36      0.17      0.24        46

    accuracy                           0.95      3595
   macro avg       0.79      0.75      0.76      3595
weighted avg       0.95      0.95      0.95      3595

The function took 2.5876 seconds to execute.

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.95      0.98      0.96      3378
     Class 1       0.82      0.93      0.87       285
     Class 2       0.99      0.87      0.93       774
     Class 3       0.32      0.11      0.16        57

    accuracy                           0.95      4494
   macro avg       0.77      0.72      0.73      4494
weighted avg       0.94      0.95      0.94      4494

The function took 0.4245 seconds to execute.
[BS=32] done in 49.9s  best-val=0.1861

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.99      0.94      0.96     10808
     Class 1       0.77      0.96      0.85       914
     Class 2       0.95      0.93      0.94      2475
     Class 3       0.35      0.92      0.51       182

    accuracy                           0.94     14379
   macro avg       0.76      0.94      0.82     14379
weighted avg       0.96      0.94      0.95     14379

The function took 0.7330 seconds to execute.

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.99      0.94      0.96      2702
     Class 1       0.79      0.93      0.85       228
     Class 2       0.93      0.95      0.94       619
     Class 3       0.33      0.91      0.49        46

    accuracy                           0.94      3595
   macro avg       0.76      0.93      0.81      3595
weighted avg       0.96      0.94      0.95      3595

The function took 0.4028 seconds to execute.

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.98      0.94      0.96      3378
     Class 1       0.78      0.94      0.85       285
     Class 2       0.93      0.93      0.93       774
     Class 3       0.35      0.82      0.49        57

    accuracy                           0.94      4494
   macro avg       0.76      0.91      0.81      4494
weighted avg       0.95      0.94      0.94      4494

The function took 0.4174 seconds to execute.
[BS=64] done in 29.0s  best-val=0.2073

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.99      0.89      0.94     10808
     Class 1       0.77      0.96      0.85       914
     Class 2       0.80      0.97      0.88      2475
     Class 3       0.31      0.88      0.46       182

    accuracy                           0.91     14379
   macro avg       0.72      0.93      0.78     14379
weighted avg       0.94      0.91      0.92     14379

# ---------------- activation sweep ---------------------------------
# Purpose:
#   Compare how different activation functions (Identity = “linear”,
#   Sigmoid, ReLU) influence performance while keeping architecture,
#   optimiser, loss, and batch size fixed (decided by the values obtained before).




BEST_BS = 32

train_loader = DataLoader(train_dataset, batch_size=BEST_BS, shuffle=True)
val_loader   = DataLoader(val_dataset,   batch_size=BEST_BS, shuffle=False)
test_loader  = DataLoader(test_dataset,  batch_size=BEST_BS, shuffle=False)

acts = {
    'linear' : nn.Identity(),
    'sigmoid': nn.Sigmoid(),
    'relu'   : nn.ReLU()
}

act_scores = {}

class ActNet(nn.Module):
    """
    Generic FFNN: the Linear layers are stored in self.layers,
    and the activation function can be selected on the fly via the `act` parameter.

    """
    def __init__(self, sizes, act):
        super().__init__()
        self.act = act

        self.layers = nn.ModuleList(
            [nn.Linear(sizes[i], sizes[i + 1]) for i in range(len(sizes) - 1)]
        )

    def forward(self, x):

        for layer in self.layers[:-1]:
            x = self.act(layer(x))
        return self.layers[-1](x)


criterion = nn.CrossEntropyLoss(weight=class_weights.to(DEVICE))

for name, act in acts.items():
    net = ActNet(ref_arch, act).to(DEVICE)
    opt = optim.AdamW(net.parameters(), lr=5e-4)

    training_loop(net, train_loader, val_loader,
                  train_dataset, val_dataset, DEVICE,
                  opt, criterion,
                  max_epochs=50, patience=15,
                  verbose_every=0, cfg_name=f"ACT={name}")

    act_scores[name] = testing_model(net, test_loader, DEVICE)


(pd.Series({k: v for k, v in act_scores.items()}))

[ACT=linear] done in 48.9s  best-val=0.4296

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.99      0.79      0.88      3378
     Class 1       0.32      0.89      0.47       285
     Class 2       0.95      0.84      0.89       774
     Class 3       0.16      0.89      0.26        57

    accuracy                           0.81      4494
   macro avg       0.60      0.85      0.63      4494
weighted avg       0.93      0.81      0.85      4494

The function took 0.3116 seconds to execute.
[ACT=sigmoid] done in 52.1s  best-val=0.3249

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.96      0.91      0.94      3378
     Class 1       0.68      0.88      0.77       285
     Class 2       0.98      0.86      0.91       774
     Class 3       0.20      0.88      0.33        57

    accuracy                           0.90      4494
   macro avg       0.71      0.88      0.74      4494
weighted avg       0.94      0.90      0.91      4494

The function took 0.3324 seconds to execute.
[ACT=relu] done in 52.9s  best-val=0.1834

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.99      0.90      0.95      3378
     Class 1       0.77      0.94      0.84       285
     Class 2       0.82      0.97      0.89       774
     Class 3       0.36      0.91      0.52        57

    accuracy                           0.92      4494
   macro avg       0.73      0.93      0.80      4494
weighted avg       0.94      0.92      0.92      4494

The function took 0.3141 seconds to execute.

# ----------------  optimiser sweep ----------------------------------
BEST_BS = 32


train_loader = DataLoader(train_dataset, batch_size=BEST_BS, shuffle=True)
val_loader   = DataLoader(val_dataset,   batch_size=BEST_BS, shuffle=False)
test_loader  = DataLoader(test_dataset,  batch_size=BEST_BS, shuffle=False)


def mk_opt(tag, net, lr, mom=None):
    if tag == 'sgd'  : return optim.SGD(net.parameters(), lr=lr)
    if tag == 'mom'  : return optim.SGD(net.parameters(), lr=lr, momentum=mom)
    if tag == 'adamw': return optim.AdamW(net.parameters(), lr=lr)


opt_cfgs = [
    ('sgd',   1e-2,  None),
    ('mom',   1e-2,  0.1),
    ('mom',   1e-2,  0.5),
    ('mom',   1e-2,  0.9),
    ('adamw', 5e-4,  None)
]


act_scores = {}
for tag, lr, m in opt_cfgs:
    net = ActNet(ref_arch, nn.ReLU()).to(DEVICE)
    opt = mk_opt(tag, net, lr, m)

    training_loop(net, train_loader, val_loader,
                  train_dataset, val_dataset, DEVICE,
                  opt, nn.CrossEntropyLoss(weight=class_weights.to(DEVICE)),
                  max_epochs=50, patience=15,
                  verbose_every=0, cfg_name=f"OPT={tag},m={m}")

    act_scores[(tag, m)] = testing_model(net, test_loader, DEVICE)

pd.Series(act_scores)

[OPT=sgd,m=None] done in 45.8s  best-val=0.2092

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.98      0.89      0.93      3378
     Class 1       0.74      0.94      0.83       285
     Class 2       0.83      0.94      0.88       774
     Class 3       0.25      0.88      0.39        57

    accuracy                           0.90      4494
   macro avg       0.70      0.91      0.76      4494
weighted avg       0.93      0.90      0.91      4494

The function took 0.3073 seconds to execute.
[OPT=mom,m=0.1] done in 47.7s  best-val=0.2165

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.99      0.88      0.93      3378
     Class 1       0.75      0.94      0.83       285
     Class 2       0.75      0.96      0.84       774
     Class 3       0.36      0.89      0.51        57

    accuracy                           0.90      4494
   macro avg       0.71      0.92      0.78      4494
weighted avg       0.92      0.90      0.90      4494

The function took 0.3161 seconds to execute.
[OPT=mom,m=0.5] done in 47.2s  best-val=0.2058

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.98      0.95      0.96      3378
     Class 1       0.78      0.94      0.85       285
     Class 2       0.94      0.91      0.93       774
     Class 3       0.37      0.86      0.52        57

    accuracy                           0.94      4494
   macro avg       0.77      0.91      0.81      4494
weighted avg       0.95      0.94      0.94      4494

The function took 0.4729 seconds to execute.
[OPT=mom,m=0.9] done in 47.4s  best-val=0.1827

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.98      0.94      0.96      3378
     Class 1       0.79      0.94      0.86       285
     Class 2       0.90      0.94      0.92       774
     Class 3       0.38      0.88      0.53        57

    accuracy                           0.94      4494
   macro avg       0.76      0.92      0.82      4494
weighted avg       0.95      0.94      0.94      4494

The function took 0.3281 seconds to execute.
[OPT=adamw,m=None] done in 53.1s  best-val=0.1795

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.98      0.94      0.96      3378
     Class 1       0.77      0.94      0.84       285
     Class 2       0.93      0.92      0.92       774
     Class 3       0.31      0.89      0.46        57

    accuracy                           0.93      4494
   macro avg       0.75      0.92      0.80      4494
weighted avg       0.95      0.93      0.94      4494

# ---------------- Task-5 · deep 6-layer grid ------------------------
# Purpose:
#   Sweep a 6-layer MLP across architecture variants and weight-decay values.
#   Variants: plain, dropout (“drop”), batch-normalisation (“bn”),
#             and batch-normalisation + dropout (“bn+drop”).
#   Weight-decay grid: 0, 1e-4, 1e-3, 1e-2.
# Procedure:
#   1. Build DataLoaders with batch size 128.
#   2. Define `DeepNet`, which can insert BatchNorm and/or Dropout after each
#      hidden Linear layer.
#   3. Helper `deep_variant` returns the requested model plus an AdamW optimiser.
#   4. For every (variant × weight-decay) pair:
#        • Train ≤50 epochs with early-stopping (patience 15) via `training_loop`.
#        • Save loss curves and accuracies (train / val / test).
#   5. Collect results into a DataFrame for quick comparison.
import itertools
import torch.nn as nn
from torch.utils.data import DataLoader


BS_DEEP = 128
train_loader128 = DataLoader(train_dataset, batch_size=BS_DEEP, shuffle=True)
val_loader128   = DataLoader(val_dataset,   batch_size=BS_DEEP, shuffle=False)
test_loader128  = DataLoader(test_dataset,  batch_size=BS_DEEP, shuffle=False)


deep_arch = [input_dim, 256, 128, 64, 32, 16, n_classes]

class DeepNet(nn.Module):
    """Deep MLP: 6 hidden layers, optional BatchNorm and/or Dropout."""
    def __init__(self, sizes, use_bn=False, use_do=False, p_drop=0.2):
        super().__init__()
        layers = []
        for i in range(len(sizes)-2):
            layers.append(nn.Linear(sizes[i], sizes[i+1]))
            if use_bn:
                layers.append(nn.BatchNorm1d(sizes[i+1]))
            layers.append(nn.ReLU())
            if use_do:
                layers.append(nn.Dropout(p_drop))
        layers.append(nn.Linear(sizes[-2], sizes[-1]))
        self.net = nn.Sequential(*layers)

    def forward(self, x):
        return self.net(x)

def deep_variant(kind:str, weight_decay:float=0.0):
     """Factory: build DeepNet variant + AdamW optimiser with given weight-decay."""
    if   kind == 'plain':
        model = DeepNet(deep_arch, use_bn=False, use_do=False)
    elif kind == 'drop':
        model = DeepNet(deep_arch, use_bn=False, use_do=True)
    elif kind == 'bn':
        model = DeepNet(deep_arch, use_bn=True,  use_do=False)
    elif kind == 'bn+drop':
        model = DeepNet(deep_arch, use_bn=True,  use_do=True)
    else:
        raise ValueError(f"kind '{kind}' ")

    model = model.to(DEVICE)
    optimizer = optim.AdamW(model.parameters(), lr=5e-4, weight_decay=weight_decay)
    return model, optimizer

kind_grid = ['plain', 'drop', 'bn', 'bn+drop']
wd_grid   = [0.0, 1e-4, 1e-3, 1e-2]

criterion = nn.CrossEntropyLoss(weight=class_weights.to(DEVICE))

deep_scores, deep_curves = {}, {}
for kd, wd in itertools.product(kind_grid, wd_grid):
    tag = f"{kd}  wd={wd:.0e}"
    model, optimizer = deep_variant(kd, weight_decay=wd)

    hist, model = training_loop(model,
                                train_loader128, val_loader128,
                                train_dataset,  val_dataset,
                                DEVICE, optimizer, criterion,
                                max_epochs=50, patience=15,
                                verbose_every=10, cfg_name=tag)

    deep_curves[tag] = hist
    deep_scores[tag] = {
        'train_acc': testing_model(model, train_loader128, DEVICE),
        'val_acc'  : testing_model(model, val_loader128,   DEVICE),
        'test_acc' : testing_model(model, test_loader128,     DEVICE)
    }

pd.DataFrame(deep_scores).T

[plain  wd=0e+00] ep 010  tr=0.2463  val=0.2872
[plain  wd=0e+00] ep 020  tr=0.2223  val=0.2236
[plain  wd=0e+00] ep 030  tr=0.2030  val=0.2171
[plain  wd=0e+00] ep 040  tr=0.1970  val=0.2095
[plain  wd=0e+00] ep 050  tr=0.1865  val=0.1961
[plain  wd=0e+00] done in 21.0s  best-val=0.1961

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.99      0.92      0.96     10808
     Class 1       0.78      0.96      0.86       914
     Class 2       0.86      0.96      0.91      2475
     Class 3       0.40      0.88      0.55       182

    accuracy                           0.93     14379
   macro avg       0.76      0.93      0.82     14379
weighted avg       0.95      0.93      0.94     14379

The function took 0.4246 seconds to execute.

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.99      0.93      0.96      2702
     Class 1       0.81      0.93      0.86       228
     Class 2       0.86      0.97      0.91       619
     Class 3       0.37      0.89      0.53        46

    accuracy                           0.93      3595
   macro avg       0.76      0.93      0.81      3595
weighted avg       0.95      0.93      0.94      3595

The function took 0.2571 seconds to execute.

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.99      0.92      0.96      3378
     Class 1       0.78      0.93      0.85       285
     Class 2       0.85      0.96      0.90       774
     Class 3       0.39      0.84      0.54        57

    accuracy                           0.93      4494
   macro avg       0.75      0.92      0.81      4494
weighted avg       0.94      0.93      0.93      4494

The function took 0.2561 seconds to execute.
[plain  wd=1e-04] ep 010  tr=0.2533  val=0.2698
[plain  wd=1e-04] ep 020  tr=0.2123  val=0.2155
[plain  wd=1e-04] ep 030  tr=0.2019  val=0.2126
[plain  wd=1e-04] ep 040  tr=0.1998  val=0.2048
[plain  wd=1e-04] ep 050  tr=0.1905  val=0.2041
[plain  wd=1e-04] done in 20.4s  best-val=0.1973

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.99      0.91      0.95     10808
     Class 1       0.77      0.96      0.86       914
     Class 2       0.86      0.91      0.88      2475
     Class 3       0.28      0.96      0.43       182

    accuracy                           0.92     14379
   macro avg       0.72      0.93      0.78     14379
weighted avg       0.94      0.92      0.92     14379

The function took 0.6328 seconds to execute.

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.99      0.91      0.95      2702
     Class 1       0.79      0.93      0.86       228
     Class 2       0.86      0.91      0.88       619
     Class 3       0.25      0.98      0.39        46

    accuracy                           0.91      3595
   macro avg       0.72      0.93      0.77      3595
weighted avg       0.95      0.91      0.93      3595

The function took 0.3776 seconds to execute.

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.98      0.92      0.95      3378
     Class 1       0.77      0.94      0.85       285
     Class 2       0.85      0.89      0.87       774
     Class 3       0.26      0.91      0.41        57

    accuracy                           0.91      4494
   macro avg       0.72      0.91      0.77      4494
weighted avg       0.94      0.91      0.92      4494

The function took 0.3903 seconds to execute.
[plain  wd=1e-03] ep 010  tr=0.2531  val=0.2488
[plain  wd=1e-03] ep 020  tr=0.2251  val=0.2401
[plain  wd=1e-03] ep 030  tr=0.2084  val=0.2294
[plain  wd=1e-03] ep 040  tr=0.2109  val=0.2266
[plain  wd=1e-03] ep 050  tr=0.1877  val=0.2031
[plain  wd=1e-03] done in 20.8s  best-val=0.1945

Classification Report:
              precision    recall  f1-score   support

     Class 0       0.99      0.90      0.94     10808
     Class 1       0.77      0.96      0.85       914
     Class 2       0.81      0.95      0.87      2475
     Class 3       0.29      0.95      0.44       182

    accuracy                           0.91     14379
   macro avg       0.72      0.94      0.78     14379
weighted avg       0.94      0.91      0.92     14379

	Flow Duration	Flow IAT Mean	Fwd PSH Flags	Bwd Packet Length Mean	Bwd Packet Length Max	Flow Bytes/s	Down/Up Ratio	SYN Flag Count	Fwd Packet Length Mean	Fwd IAT Std	Packet Length Mean	Fwd Packet Length Max	Subflow Fwd Packets	Flow Packets/s	Total Fwd Packets	Destination Port
count	3.150700e+04	3.150700e+04	31507.000000	31507.000000	31507.000000	3.148700e+04	31507.000000	31507.000000	31507.000000	3.150700e+04	31507.000000	31507.000000	31507.000000	3.150700e+04	31507.000000	31507.000000
mean	1.647581e+07	1.438258e+06	0.038277	310.232250	896.532040	inf	0.639382	0.038277	44.150959	4.653801e+06	163.155706	170.496652	6.181293	inf	6.181293	7367.544292
std	3.526316e+07	4.122202e+06	0.191868	605.892466	1945.454241	NaN	0.531334	0.191868	115.215481	1.185644e+07	290.189858	501.331438	109.450416	NaN	109.450416	17104.518207
min	-1.000000e+00	-1.000000e+00	0.000000	0.000000	0.000000	-1.200000e+07	0.000000	0.000000	0.000000	0.000000e+00	0.000000	0.000000	1.000000	-2.000000e+06	1.000000	0.000000
25%	6.500000e+01	5.500000e+01	0.000000	5.500000	6.000000	1.199456e+02	0.000000	0.000000	2.000000	0.000000e+00	3.333333	2.000000	1.000000	2.032897e+00	1.000000	53.000000
50%	2.427100e+04	8.096333e+03	0.000000	51.000000	56.000000	4.830598e+03	1.000000	0.000000	32.000000	0.000000e+00	47.800000	35.000000	2.000000	1.581044e+02	2.000000	80.000000
75%	5.216842e+06	5.558727e+05	0.000000	176.000000	271.000000	1.538462e+05	1.000000	0.000000	50.000000	7.627202e+05	112.000000	198.000000	5.000000	3.174603e+04	5.000000	1066.000000
max	1.199985e+08	1.140000e+08	1.000000	3865.000000	11595.000000	inf	10.000000	1.000000	3431.806400	8.280000e+07	1903.272700	23360.000000	18605.000000	inf	18605.000000	65533.000000

	Flow Duration	Flow IAT Mean	Fwd PSH Flags	Bwd Packet Length Mean	Bwd Packet Length Max	Flow Bytes/s	Down/Up Ratio	SYN Flag Count	Fwd Packet Length Mean	Fwd IAT Std	Packet Length Mean	Fwd Packet Length Max	Subflow Fwd Packets	Flow Packets/s	Total Fwd Packets	Destination Port
Label
Benign	1.131130e+07	9.482337e+05	0.0603	165.798787	411.053000	inf	0.650750	0.0603	57.110944	1.220772e+06	107.074714	206.50040	7.244800	inf	7.244800	9408.6374
Brute Force	6.506182e+06	1.546228e+06	0.0000	41.622306	91.857332	1.792807e+02	0.142668	0.0000	16.869312	3.264638e+06	24.509333	54.50564	12.371599	3684.902275	12.371599	80.0000
DoS Hulk	5.653013e+07	4.776383e+06	0.0000	1272.397426	3969.474000	3.213723e+04	0.388800	0.0000	43.682147	2.345695e+07	589.297444	230.89800	5.229000	inf	5.229000	80.0000
PortScan	8.436853e+04	2.768802e+04	0.0000	6.759967	8.035200	inf	0.994200	0.0000	1.002520	1.465057e+03	3.125954	1.04000	1.013800	inf	1.013800	8687.1820

	Flow Duration								Flow IAT Mean		...	Total Fwd Packets		Destination Port
	count	mean	std	min	25%	50%	75%	max	count	mean	...	75%	max	count	mean	std	min	25%	50%	75%	max
Label
Benign	20000.0	1.131130e+07	3.030644e+07	-1.0	182.0	31112.5	429108.00	119998546.0	20000.0	9.482337e+05	...	4.0	18605.0	20000.0	9408.6374	19840.517250	0.0	53.00	80.0	443.0	65533.0
Brute Force	1507.0	6.506182e+06	6.415511e+06	20.0	5180548.0	5552319.0	5838340.00	35451960.0	1507.0	1.546228e+06	...	3.0	204.0	1507.0	80.0000	0.000000	80.0	80.00	80.0	80.0	80.0
DoS Hulk	5000.0	5.653013e+07	4.621738e+07	0.0	986.0	84904369.0	98619498.25	119165532.0	5000.0	4.776383e+06	...	7.0	13.0	5000.0	80.0000	0.000000	80.0	80.00	80.0	80.0	80.0
PortScan	5000.0	8.436853e+04	2.014134e+06	0.0	42.0	47.0	61.00	77038552.0	5000.0	2.768802e+04	...	1.0	6.0	5000.0	8687.1820	13472.839374	1.0	1123.75	3690.0	8291.0	65389.0

	Flow Duration	Flow IAT Mean	Fwd PSH Flags	Bwd Packet Length Mean	Bwd Packet Length Max	Flow Bytes/s	Down/Up Ratio	SYN Flag Count	Fwd Packet Length Mean	Fwd IAT Std	Packet Length Mean	Fwd Packet Length Max	Subflow Fwd Packets	Flow Packets/s	Total Fwd Packets	Destination Port	Label
0	303376	3.033760e+04	0	749.4	1448	12743.263	0	0	19.833334	98776.15	322.16666	119	6	36.258636	6	443	0
1	117	1.170000e+02	0	0.0	0	0.000	1	0	0.000000	0.00	0.00000	0	1	17094.018000	1	52631	0
2	142	1.420000e+02	0	0.0	0	84507.040	0	0	6.000000	0.00	6.00000	6	2	14084.507000	2	80	0
3	191	6.366667e+01	0	52.0	52	921465.940	1	0	36.000000	0.00	42.40000	36	2	20942.408000	2	53	0
4	4	4.000000e+00	0	0.0	0	3000000.000	0	0	6.000000	0.00	6.00000	6	2	500000.000000	2	49467	0
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
31502	5710955	1.903652e+06	0	0.0	0	0.000	0	0	0.000000	4037277.80	0.00000	0	3	0.700408	3	80	1
31503	5862561	1.954187e+06	0	0.0	0	0.000	0	0	0.000000	4144374.80	0.00000	0	3	0.682296	3	80	1
31504	5854872	1.951624e+06	0	0.0	0	0.000	0	0	0.000000	4139029.80	0.00000	0	3	0.683192	3	80	1
31505	5555547	1.851849e+06	0	0.0	0	0.000	0	0	0.000000	3927356.50	0.00000	0	3	0.720001	3	80	1
31506	5018204	1.672735e+06	0	0.0	0	0.000	0	0	0.000000	3547404.80	0.00000	0	3	0.797098	3	80	1

	Flow Duration	Flow IAT Mean	Fwd PSH Flags	Bwd Packet Length Mean	Bwd Packet Length Max	Flow Bytes/s	Down/Up Ratio	SYN Flag Count	Fwd Packet Length Mean	Fwd IAT Std	Packet Length Mean	Fwd Packet Length Max	Subflow Fwd Packets	Flow Packets/s	Total Fwd Packets	Destination Port	Label
count	3.150700e+04	3.150700e+04	31507.000000	31507.000000	31507.000000	3.148700e+04	31507.000000	31507.000000	31507.000000	3.150700e+04	31507.000000	31507.000000	31507.000000	3.150700e+04	31507.000000	31507.000000	31507.000000
mean	1.647581e+07	1.438258e+06	0.038277	310.232250	896.532040	inf	0.639382	0.038277	44.150959	4.653801e+06	163.155706	170.496652	6.181293	inf	6.181293	7367.544292	0.841305
std	3.526316e+07	4.122202e+06	0.191868	605.892466	1945.454241	NaN	0.531334	0.191868	115.215481	1.185644e+07	290.189858	501.331438	109.450416	NaN	109.450416	17104.518207	1.184531
min	-1.000000e+00	-1.000000e+00	0.000000	0.000000	0.000000	-1.200000e+07	0.000000	0.000000	0.000000	0.000000e+00	0.000000	0.000000	1.000000	-2.000000e+06	1.000000	0.000000	0.000000
25%	6.500000e+01	5.500000e+01	0.000000	5.500000	6.000000	1.199456e+02	0.000000	0.000000	2.000000	0.000000e+00	3.333333	2.000000	1.000000	2.032897e+00	1.000000	53.000000	0.000000
50%	2.427100e+04	8.096333e+03	0.000000	51.000000	56.000000	4.830598e+03	1.000000	0.000000	32.000000	0.000000e+00	47.800000	35.000000	2.000000	1.581044e+02	2.000000	80.000000	0.000000
75%	5.216842e+06	5.558727e+05	0.000000	176.000000	271.000000	1.538462e+05	1.000000	0.000000	50.000000	7.627202e+05	112.000000	198.000000	5.000000	3.174603e+04	5.000000	1066.000000	2.000000
max	1.199985e+08	1.140000e+08	1.000000	3865.000000	11595.000000	inf	10.000000	1.000000	3431.806400	8.280000e+07	1903.272700	23360.000000	18605.000000	inf	18605.000000	65533.000000	3.000000

LAB 1¶

Data-preprocessing¶

infinite values analysis¶

null values¶

label counts¶

Label Encoder¶

drop phase¶

check for negative values¶

drop duplicates¶

Plots analysis¶

Correlation Matrix¶

skewness analysis¶

Dataset split¶

Analysis for features (raw data - not scaled)¶

Outliers analysis (raw data)¶

Standard Scaler¶

Outliers analysis¶

Analysis for feature¶

PCA + T-SNE¶

DataLoader ( SS )¶

Task 2: Shallow Neural Network ( Log + Robust)¶

Hyperparameter¶

h:32¶

h: 64¶

h: 128¶

ReLu activation¶

Task 3 : Impact of specific features¶

Dest-port dropping¶

class-weight¶

Task 4: Deep NN¶

Builder¶

best - batch-size¶

best - activation function¶

best - optimizer¶

Task 5: Overfitting and regularization¶

	Flow Duration	Flow IAT Mean	Bwd Packet Length Mean	Bwd Packet Length Max	Flow Bytes/s	Down/Up Ratio	Fwd Packet Length Mean	Fwd IAT Std	Packet Length Mean	Fwd Packet Length Max	Subflow Fwd Packets	Flow Packets/s	Total Fwd Packets	Destination Port	Label
0	303376	30337.600000	749.4	1448	12743.263	0	19.833334	98776.15	322.16666	119	6	36.258636	6	443	Benign
1	117	117.000000	0.0	0	0.000	1	0.000000	0.00	0.00000	0	1	17094.018000	1	52631	Benign
2	142	142.000000	0.0	0	84507.040	0	6.000000	0.00	6.00000	6	2	14084.507000	2	80	Benign
3	191	63.666668	52.0	52	921465.940	1	36.000000	0.00	42.40000	36	2	20942.408000	2	53	Benign
4	4	4.000000	0.0	0	3000000.000	0	6.000000	0.00	6.00000	6	2	500000.000000	2	49467	Benign

	Flow Bytes/s	Subflow Fwd Packets	Flow Packets/s	Total Fwd Packets	Destination Port	Label
2130	NaN	2	inf	2	52606	0
16361	NaN	2	inf	2	19152	0
20043	NaN	2	inf	2	80	2
20524	NaN	2	inf	2	80	2
20650	NaN	2	inf	2	80	2
20779	NaN	2	inf	2	80	2
20839	NaN	2	inf	2	80	2
21003	NaN	2	inf	2	80	2
21449	NaN	2	inf	2	80	2
21700	NaN	2	inf	2	80	2
22152	NaN	2	inf	2	80	2
22955	NaN	2	inf	2	80	2
23687	NaN	2	inf	2	80	2
24006	NaN	2	inf	2	80	2
24045	NaN	2	inf	2	80	2
24404	NaN	2	inf	2	80	2
24452	NaN	2	inf	2	80	2
24454	NaN	2	inf	2	80	2
24598	NaN	2	inf	2	80	2
24608	NaN	2	inf	2	80	2

	Flow Duration	Flow IAT Mean	Fwd PSH Flags	Bwd Packet Length Mean	Bwd Packet Length Max	Flow Bytes/s	Down/Up Ratio	SYN Flag Count	Fwd Packet Length Mean	Fwd IAT Std	Packet Length Mean	Fwd Packet Length Max	Subflow Fwd Packets	Flow Packets/s	Total Fwd Packets	Destination Port	Label
12379	-1	-1.0	0	6.0	6	-12000000.0	1	0	6.0	0.0	6.0	6	1	-2000000.0	1	443	0
15783	-1	-1.0	0	6.0	6	-12000000.0	1	0	6.0	0.0	6.0	6	1	-2000000.0	1	80	0

	Flow Duration	Flow IAT Mean	Fwd PSH Flags	Bwd Packet Length Mean	Bwd Packet Length Max	Flow Bytes/s	Down/Up Ratio	SYN Flag Count	Fwd Packet Length Mean	Fwd IAT Std	Packet Length Mean	Fwd Packet Length Max	Subflow Fwd Packets	Flow Packets/s	Total Fwd Packets	Destination Port
Label
0	750	750	750	750	750	750	750	750	750	750	750	750	750	750	750	750
1	80	80	80	80	80	80	80	80	80	80	80	80	80	80	80	80
2	1114	1114	1114	1114	1114	1114	1114	1114	1114	1114	1114	1114	1114	1114	1114	1114
3	150	150	150	150	150	150	150	150	150	150	150	150	150	150	150	150

	Flow Duration	Flow IAT Mean	Fwd PSH Flags	Bwd Packet Length Mean	Bwd Packet Length Max	Flow Bytes/s	Down/Up Ratio	SYN Flag Count	Fwd Packet Length Mean	Fwd IAT Std	Packet Length Mean	Fwd Packet Length Max	Subflow Fwd Packets	Flow Packets/s	Total Fwd Packets	Destination Port	Label
count	2.938400e+04	2.938400e+04	29384.000000	29384.000000	29384.000000	2.938400e+04	29384.000000	29384.000000	29384.000000	2.938400e+04	29384.000000	29384.000000	29384.000000	2.938400e+04	29384.000000	29384.000000	29384.000000
mean	1.766618e+07	1.542164e+06	0.039375	331.783641	960.443711	1.441052e+06	0.665498	0.039375	46.791664	4.990040e+06	174.207776	182.240199	6.487748	4.897875e+04	6.487748	7834.983903	0.806902
std	3.622577e+07	4.249715e+06	0.194489	621.821967	1999.389155	2.553726e+07	0.527022	0.194489	118.808422	1.220878e+07	297.407429	517.136534	113.329113	1.942472e+05	113.329113	17570.738668	1.187124
min	1.000000e+00	5.000000e-01	0.000000	0.000000	0.000000	0.000000e+00	0.000000	0.000000	0.000000	0.000000e+00	0.000000	0.000000	1.000000	1.748631e-02	1.000000	0.000000	0.000000
25%	8.400000e+01	6.133333e+01	0.000000	6.000000	6.000000	1.212895e+02	0.000000	0.000000	2.000000	0.000000e+00	3.333333	2.000000	1.000000	1.279737e+00	1.000000	53.000000	0.000000
50%	3.111750e+04	1.113595e+04	0.000000	63.000000	69.000000	5.130109e+03	1.000000	0.000000	34.000000	0.000000e+00	52.333332	37.000000	2.000000	1.146816e+02	2.000000	80.000000	0.000000
75%	5.490093e+06	9.073129e+05	0.000000	202.593750	429.000000	1.481481e+05	1.000000	0.000000	51.000000	1.959710e+06	128.333330	265.250000	5.000000	2.666667e+04	5.000000	1216.250000	2.000000
max	1.199985e+08	1.140000e+08	1.000000	3865.000000	11595.000000	2.070000e+09	10.000000	1.000000	3431.806400	8.280000e+07	1903.272700	23360.000000	18605.000000	3.000000e+06	18605.000000	65533.000000	3.000000

	Flow Duration	Flow IAT Mean	Fwd PSH Flags	Bwd Packet Length Mean	Bwd Packet Length Max	Flow Bytes/s	Down/Up Ratio	SYN Flag Count	Fwd Packet Length Mean	Fwd IAT Std	Packet Length Mean	Fwd Packet Length Max	Subflow Fwd Packets	Flow Packets/s	Total Fwd Packets	Destination Port	Label
24972	44	44.0	0	6.0	6	136363.640	1	0	0.0	0.0	2.000000	0	1	45454.547	1	84	3
24973	54	54.0	0	6.0	6	111111.110	1	0	0.0	0.0	2.000000	0	1	37037.035	1	4449	3
24974	27	27.0	0	6.0	6	222222.220	1	0	0.0	0.0	2.000000	0	1	74074.070	1	12345	3
24975	52	52.0	0	6.0	6	153846.160	1	0	2.0	0.0	3.333333	2	1	38461.540	1	4125	3
24976	32	32.0	0	6.0	6	187500.000	1	0	0.0	0.0	2.000000	0	1	62500.000	1	1984	3
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
29966	44	44.0	0	6.0	6	136363.640	1	0	0.0	0.0	2.000000	0	1	45454.547	1	32	3
29967	61	61.0	0	6.0	6	131147.550	1	0	2.0	0.0	3.333333	2	1	32786.887	1	1028	3
29968	42	42.0	0	6.0	6	190476.190	1	0	2.0	0.0	3.333333	2	1	47619.047	1	28201	3
29969	91	91.0	0	6.0	6	87912.086	1	0	2.0	0.0	3.333333	2	1	21978.021	1	7937	3
29970	92	92.0	0	6.0	6	65217.390	1	0	0.0	0.0	2.000000	0	1	21739.130	1	25	3

	train	val	test
ARCH#001-[32, 32]	{'accuracy': 93.18450518116698, 'precision': [...	{'accuracy': 92.87899860917942, 'precision': [...	{'accuracy': 93.12416555407209, 'precision': [...
ARCH#002-[32, 16]	{'accuracy': 90.88253703317338, 'precision': [...	{'accuracy': 91.04311543810849, 'precision': [...	{'accuracy': 90.96573208722741, 'precision': [...
ARCH#003-[32, 8]	{'accuracy': 91.18853884136588, 'precision': [...	{'accuracy': 91.23783031988873, 'precision': [...	{'accuracy': 91.4997774810859, 'precision': [0...
ARCH#004-[32, 4]	{'accuracy': 92.00222546769594, 'precision': [...	{'accuracy': 91.84979137691238, 'precision': [...	{'accuracy': 91.90031152647975, 'precision': [...
ARCH#005-[32, 2]	{'accuracy': 90.22185131093956, 'precision': [...	{'accuracy': 90.34770514603616, 'precision': [...	{'accuracy': 89.56386292834891, 'precision': [...
...	...	...	...
ARCH#242-[4, 4, 4, 4, 2]	{'accuracy': 82.80130746227137, 'precision': [...	{'accuracy': 82.89290681502086, 'precision': [...	{'accuracy': 83.31108144192257, 'precision': [...
ARCH#243-[4, 4, 4, 2, 2]	{'accuracy': 42.116976145768135, 'precision': ...	{'accuracy': 41.55771905424201, 'precision': [...	{'accuracy': 41.477525589675125, 'precision': ...
ARCH#244-[4, 4, 2, 2, 2]	{'accuracy': 75.16517143055846, 'precision': [...	{'accuracy': 75.15994436717664, 'precision': [...	{'accuracy': 75.16688918558077, 'precision': [...
ARCH#245-[4, 2, 2, 2, 2]	{'accuracy': 75.16517143055846, 'precision': [...	{'accuracy': 75.15994436717664, 'precision': [...	{'accuracy': 75.10013351134846, 'precision': [...
ARCH#246-[2, 2, 2, 2, 2]	{'accuracy': 75.16517143055846, 'precision': [...	{'accuracy': 75.15994436717664, 'precision': [...	{'accuracy': 75.16688918558077, 'precision': [...

	train_acc	val_acc	test_acc
1	{'accuracy': 95.01356144377216, 'precision': [...	{'accuracy': 95.35465924895689, 'precision': [...	{'accuracy': 94.57053849577214, 'precision': [...
32	{'accuracy': 94.13728353849363, 'precision': [...	{'accuracy': 94.1029207232267, 'precision': [0...	{'accuracy': 93.92523364485982, 'precision': [...
64	{'accuracy': 91.13290214896725, 'precision': [...	{'accuracy': 91.15438108484005, 'precision': [...	{'accuracy': 91.18825100133512, 'precision': [...
128	{'accuracy': 91.08422004311844, 'precision': [...	{'accuracy': 90.93184979137692, 'precision': [...	{'accuracy': 91.3440142412105, 'precision': [0...
512	{'accuracy': 86.75846720912442, 'precision': [...	{'accuracy': 86.98191933240612, 'precision': [...	{'accuracy': 86.55985758789497, 'precision': [...

	0
linear	{'accuracy': 80.77436582109479, 'precision': [...
sigmoid	{'accuracy': 89.942145082332, 'precision': [0....
relu	{'accuracy': 91.85580774365822, 'precision': [...

		0
sgd	NaN	{'accuracy': 90.09790832220739, 'precision': [...
mom	0.1	{'accuracy': 89.76412995104583, 'precision': [...
	0.5	{'accuracy': 93.81397418780595, 'precision': [...
	0.9	{'accuracy': 93.65821094793057, 'precision': [...
adamw	NaN	{'accuracy': 93.25767690253672, 'precision': [...

Configuration	Best Val Loss	Val Macro-F1	Test Macro-F1
Plain, wd=0	0.1961	0.82	0.81
Plain, wd=1e-4	0.1973	0.78	0.77
Plain, wd=1e-3	0.1945	0.78	0.77
Plain, wd=1e-2	0.1927	0.79	0.80
Dropout, wd=0	0.2136	0.78	0.79
Dropout, wd=1e-4	0.2108	0.75	0.74
Dropout, wd=1e-3	0.2077	0.76	0.75
Dropout, wd=1e-2	0.2139	0.80	0.80
BatchNorm, wd=0	0.2052	0.79	0.79
BatchNorm, wd=1e-4	0.2032	0.80	0.80
BatchNorm, wd=1e-3	0.1981	0.77	0.76
BatchNorm, wd=1e-2	0.2241	0.82	0.81
BatchNorm + Dropout, wd=0	0.2182	0.77	0.77
BatchNorm + Dropout, wd=1e-4	0.2240	0.74	0.74
BatchNorm + Dropout, wd=1e-3	0.2145	0.80	0.79
BatchNorm + Dropout, wd=1e-2	0.2199	0.79	0.79

	train_acc	val_acc	test_acc
plain wd=0e+00	{'accuracy': 93.17755059461716, 'precision': [...	{'accuracy': 93.3240611961057, 'precision': [0...	{'accuracy': 93.03515798842902, 'precision': [...
plain wd=1e-04	{'accuracy': 91.55017734195702, 'precision': [...	{'accuracy': 91.43254520166899, 'precision': [...	{'accuracy': 91.23275478415665, 'precision': [...
plain wd=1e-03	{'accuracy': 91.00076500452047, 'precision': [...	{'accuracy': 91.21001390820584, 'precision': [...	{'accuracy': 90.85447263017356, 'precision': [...
plain wd=1e-02	{'accuracy': 92.32909103553794, 'precision': [...	{'accuracy': 92.26703755215577, 'precision': [...	{'accuracy': 92.25634178905207, 'precision': [...
drop wd=0e+00	{'accuracy': 90.04798664719382, 'precision': [...	{'accuracy': 90.1529902642559, 'precision': [0...	{'accuracy': 90.12016021361816, 'precision': [...
drop wd=1e-04	{'accuracy': 89.70721190625217, 'precision': [...	{'accuracy': 89.87482614742697, 'precision': [...	{'accuracy': 89.36359590565198, 'precision': [...
drop wd=1e-03	{'accuracy': 90.02016830099451, 'precision': [...	{'accuracy': 89.8191933240612, 'precision': [0...	{'accuracy': 89.7196261682243, 'precision': [0...
drop wd=1e-02	{'accuracy': 91.84922456359969, 'precision': [...	{'accuracy': 91.90542420027816, 'precision': [...	{'accuracy': 91.70004450378282, 'precision': [...
bn wd=0e+00	{'accuracy': 92.24563599693998, 'precision': [...	{'accuracy': 92.35048678720446, 'precision': [...	{'accuracy': 91.96706720071207, 'precision': [...
bn wd=1e-04	{'accuracy': 93.49746157590931, 'precision': [...	{'accuracy': 93.29624478442281, 'precision': [...	{'accuracy': 93.36893635959056, 'precision': [...
bn wd=1e-03	{'accuracy': 91.79358787120105, 'precision': [...	{'accuracy': 92.1557719054242, 'precision': [0...	{'accuracy': 91.45527369826435, 'precision': [...
bn wd=1e-02	{'accuracy': 93.7060991724042, 'precision': [0...	{'accuracy': 93.65785813630042, 'precision': [...	{'accuracy': 93.43569203382287, 'precision': [...
bn+drop wd=0e+00	{'accuracy': 90.92426455247235, 'precision': [...	{'accuracy': 91.04311543810849, 'precision': [...	{'accuracy': 90.78771695594125, 'precision': [...
bn+drop wd=1e-04	{'accuracy': 89.36643716531051, 'precision': [...	{'accuracy': 89.59666203059805, 'precision': [...	{'accuracy': 89.02981753449043, 'precision': [...
bn+drop wd=1e-03	{'accuracy': 91.78663328465123, 'precision': [...	{'accuracy': 91.93324061196105, 'precision': [...	{'accuracy': 91.58878504672897, 'precision': [...
bn+drop wd=1e-02	{'accuracy': 91.8979066694485, 'precision': [0...	{'accuracy': 92.04450625869262, 'precision': [...	{'accuracy': 91.43302180685359, 'precision': [...